Tygart Media

Author: Will Tygart

  • Why AI Agents Are Different From Chatbots, Automations, and APIs

    Why AI Agents Are Different From Chatbots, Automations, and APIs

    These terms get used interchangeably. They’re not the same thing. Here’s the actual distinction between each one, where the lines get genuinely blurry, and which category fits what you’re actually trying to build.

    Chatbots

    A chatbot is a software interface designed to simulate conversation. The defining characteristic: it’s stateless and reactive. You send a message; it responds; the exchange is complete. Each interaction is largely independent.

    Traditional chatbots (pre-LLM) operated on decision trees — “if the user says X, respond with Y.” Modern LLM-powered chatbots use language models to generate responses, which makes them dramatically more capable and flexible — but the fundamental architecture is the same: you ask, it answers, you ask again.

    What chatbots are good at: answering questions, providing information, routing conversations, handling defined service scenarios with natural language flexibility. What they’re not: action-takers. A chatbot can tell you how to cancel your subscription. An agent can cancel it.

    Automations

    Automations are rule-based workflows that execute when triggered. Zapier, Make, and similar tools are the canonical examples. When event A happens, do B, then C, then D.

    The key characteristic: the path is predefined. Every step is specified by the person who built the automation. If an unexpected situation arises that the automation wasn’t built for, it either fails or skips the step. There’s no reasoning about what to do — there’s only executing the specified path or not.

    Automations are highly reliable for well-defined, stable processes. They break when edge cases arise that weren’t anticipated. They scale perfectly for the exact task they were built for; they don’t generalize.

    APIs

    An API (Application Programming Interface) is a communication contract — a defined way for software systems to talk to each other. APIs are infrastructure, not agents or automations. They’re the mechanism through which agents and automations take action in external systems.

    When an AI agent “uses Slack,” it’s calling Slack’s API. When an automation “posts to Twitter,” it’s calling Twitter’s API. The API is the door; agents and automations are the things that open it.

    Conflating APIs with agents is a category error. An API is a tool, not a behavior pattern.

    AI Agents

    An AI agent takes a goal and figures out how to accomplish it, using tools available to it, handling unexpected situations along the way, without a human specifying each step.

    The distinguishing characteristics versus the above:

    • vs. Chatbots: Agents take action in the world; chatbots respond to messages. An agent can book the flight, not just tell you how to book it.
    • vs. Automations: Agents reason about what to do next; automations execute predefined paths. When an unexpected situation arises, an agent adapts; an automation fails or skips.
    • vs. APIs: APIs are tools an agent uses; they’re not the agent itself. The agent is the reasoning layer that decides which API to call and what to do with the result.

    Where the Lines Actually Blur

    In practice, real systems often combine these categories:

    LLM-powered chatbots with tool access: A customer service chatbot that can look up your order status, initiate a return, and send a confirmation email is starting to look like an agent — it’s taking actions, not just responding. The boundary between “advanced chatbot” and “limited agent” is genuinely fuzzy.

    Automations with AI decision steps: A Zapier workflow with an OpenAI or Claude step in the middle isn’t purely rule-based anymore — the AI step can produce variable outputs that affect what the automation does next. This is a hybrid: mostly automation, partly agentic.

    Agents with constrained scopes: An agent restricted to a single tool and a narrow task class starts to look like a sophisticated automation. The more constrained the scope, the more the distinction collapses in practice.

    The useful question isn’t “what category is this?” but “is this system reasoning about what to do, or executing a predefined path?” That’s the actual distinction that matters for how you build, monitor, and trust it.

    Why the Distinction Matters Operationally

    Reliability profile: Automations fail predictably — when an edge case hits a path that wasn’t built. Agents fail unpredictably — when their reasoning goes wrong in a way you didn’t anticipate. Different failure modes require different monitoring approaches.

    Maintenance overhead: Automations require explicit updates when processes change. Agents adapt to process changes automatically — but may adapt in unexpected ways that need to be caught and corrected.

    Auditability: Automations are fully auditable — you can read the workflow and know exactly what it does. Agents are less auditable — you can inspect their actions, but not fully predict them in advance. For compliance-sensitive contexts, this matters significantly.

    Build cost: Automations are faster to build for well-defined, stable processes. Agents are faster to deploy when the process is complex, variable, or not fully specified — because you’re specifying a goal rather than a procedure.

    For what agents can actually do in production: What AI Agents Actually Do. For a business owner’s introduction: AI Agents Explained for Business Owners. For hosted agent infrastructure: Claude Managed Agents FAQ.


    Hosted agent infrastructure pricing: Claude Managed Agents Pricing Reference.

  • What AI Agents Actually Do (Not the Hype Version)

    What AI Agents Actually Do (Not the Hype Version)

    Not the version where AI agents are going to replace all human jobs by 2030. The actual version, right now, based on what’s deployed in production.

    The Actual Definition

    What an AI agent is

    Software that takes a goal, breaks it into steps, uses tools to execute those steps, handles errors along the way, and keeps working without you directing every action. The distinguishing characteristic is autonomous multi-step execution — not just answering a question, but completing a task.

    The Key Distinction: One-Shot vs. Agentic

    Most people’s experience with AI is one-shot: you type something, the AI responds, the exchange is complete. That’s a language model doing inference. An AI agent is different in one specific way: it takes actions, checks results, and takes more actions based on what it found — often dozens of steps — without you approving each one.

    Example of one-shot AI: “Summarize this document.” You paste the document, the AI returns a summary. Done.

    Example of an AI agent doing the same task: “Research this topic and produce a summary with verified sources.” The agent searches the web, reads multiple pages, identifies conflicts between sources, runs additional searches to resolve them, synthesizes findings, and returns a summary with citations — without you specifying each search query or each page to read. You gave it a goal; it handled the steps.

    What Agents Can Actually Do

    The tools an agent can use define its capability surface. Common tool categories in production agents:

    • Web search: Query search engines and retrieve current information
    • Code execution: Write and run code in a sandboxed environment, use results to inform next steps
    • File operations: Read, write, and modify files — documents, spreadsheets, data files
    • API calls: Interact with external services — CRMs, databases, project management tools, communication platforms
    • Browser control: Navigate web pages, fill forms, extract information
    • Memory: Store and retrieve information across steps within a session, sometimes across sessions

    The combination of these tools is what makes agents capable of genuinely autonomous work. An agent that can search, write code, execute it, check the results, and write findings to a document can complete a research and analysis task that would otherwise require hours of human work — without you steering each step.

    What “Autonomous” Actually Means in Practice

    Autonomous doesn’t mean unsupervised indefinitely. Production agents are typically configured with:

    • Defined scope: The tools the agent can use, the systems it can access, the actions it’s allowed to take
    • Guardrails: Actions that require human confirmation before proceeding — making a payment, sending an email externally, modifying a production database
    • Reporting: Checkpoints where the agent surfaces what it’s done and asks whether to continue

    Autonomy is a dial, not a switch. You set how much the agent handles independently versus checks in. Most production deployments start more supervised and reduce oversight as trust in the agent’s behavior is established.

    Real Production Examples (Not Hypotheticals)

    Concrete examples from confirmed public deployments as of April 2026:

    • Rakuten: Deployed five enterprise Claude agents in one week on Anthropic’s Managed Agents platform — handling tasks across their e-commerce operations including data processing, content tasks, and operational workflows
    • Notion: Background agents that autonomously update workspace pages, synthesize database content, and process meeting notes into structured summaries without manual triggers
    • Sentry: Agents integrated into developer workflows — monitoring error streams, triaging issues, and surfacing relevant context to engineers
    • Asana: Project management agents that update task statuses, synthesize project health, and move work items based on defined triggers

    These are not pilots. These are production systems handling real operational load.

    How They’re Built

    An agent is built from three components:

    1. A language model: The reasoning layer — the part that decides what to do next, interprets tool results, and determines when the task is complete
    2. Tools: The action layer — APIs, code execution environments, file systems, or anything else the model can call to take action in the world
    3. Orchestration: The loop that connects them — manages the sequence of model calls and tool executions, maintains state between steps, handles errors

    Historically, builders had to construct the orchestration layer themselves — a significant engineering investment. Hosted platforms like Claude Managed Agents handle the orchestration layer, letting builders focus on defining the agent’s goals, tools, and guardrails rather than the mechanics of running the loop.

    What Agents Are Not Good At (Yet)

    Honest calibration on current limitations:

    • Long-horizon planning with many unknowns: Agents perform best on tasks with relatively defined scope. Open-ended exploratory work over many days with fundamentally uncertain requirements is still better handled by humans in the loop at each major decision point.
    • Tasks requiring physical world interaction: No production general-purpose physical agent exists. Software agents operating through APIs and interfaces are the current state.
    • Tasks where errors are catastrophic: Agents make mistakes. For any irreversible, high-stakes action — financial transactions, production data modifications, external communications to important relationships — human confirmation steps should remain in the loop.

    For how hosted agent infrastructure works: Claude Managed Agents FAQ. For the difference between agents and chatbots: AI Agents vs. Chatbots, Automations, and APIs. For an SMB-focused explanation: AI Agents Explained for Business Owners.


    For pricing specifics on hosted agent infrastructure: Claude Managed Agents Complete Pricing Reference.

  • How to Write Content That AI Systems Actually Cite

    How to Write Content That AI Systems Actually Cite

    Tygart Media / Content Strategy
    The Practitioner JournalField Notes
    By Will Tygart
    · Practitioner-grade
    · From the workbench

    Being cited by AI systems is not luck and it’s not purely a domain authority game. There are structural characteristics of content that make AI systems more or less likely to pull from it. Here’s what those characteristics are and how to build them in deliberately.

    Why Content Structure Determines Citation Likelihood

    AI systems — whether Perplexity, ChatGPT with web search, or Google AI Overviews — are trying to answer a question. When they search the web and retrieve candidate content, they’re looking for the passage or page that most directly and reliably answers the query. The content that wins is the content that makes the answer easiest to extract.

    This has direct structural implications. A 3,000-word narrative essay that eventually answers a question on page 2 loses to a 600-word page that answers the question in the first paragraph, provides supporting evidence, and includes a definition. Not because shorter is better, but because clarity of answer placement is better.

    The Structural Characteristics That Drive Citation

    1. Direct Answer in the First 100 Words

    Every piece of content you want AI systems to cite should answer the primary question it’s targeting before the first scroll. AI retrieval systems don’t read like humans — they identify the most relevant passage, and that passage needs to contain the answer, not just lead toward it.

    Test: take your target query and your first 100 words. Does the answer exist in those 100 words? If not, restructure until it does. The rest of the piece can develop nuance, context, and supporting evidence — but the answer must be front-loaded.

    2. Explicit Q&A Formatting

    Question-and-answer structure signals to AI systems that the content is explicitly organized around answering queries. H3 headers phrased as questions, followed by direct answers, are one of the most reliable patterns for citation capture.

    This is why FAQ sections work — not because of FAQPage schema specifically, but because the underlying structure gives AI systems a clean extraction target. Schema reinforces it; the structure is the foundation.

    3. Defined Terms and Named Concepts

    Content that defines terms clearly — “X is Y” statements — becomes citable for queries looking for definitions. AI systems frequently answer “what is X” queries by pulling the clearest definition they can find. If your content doesn’t include a crisp definitional sentence, it’s not competing for definition queries even if you’ve written a thorough treatment of the topic.

    Add definition boxes. State “AI citation rate is the percentage of sampled AI queries where your domain appears as a cited source.” Don’t bury the definition in the third paragraph of an explanation.

    4. Specific, Verifiable Facts

    AI systems weight specificity. “$0.08 per session-hour” gets cited. “A relatively modest fee” does not. “60 requests per minute for create endpoints” gets cited. “Limited rate limits apply” does not.

    Replace hedged language with concrete numbers and specific claims wherever your content supports it. Don’t fabricate specificity — wrong specific numbers are worse than honest hedging. But wherever you have real, verifiable data, make it explicit and prominent.

    5. Entity Clarity

    Content that makes clear who is speaking, what organization they represent, and what their basis for authority is gets cited more reliably. This is the E-E-A-T signal applied to AI citation: the system needs to assess whether this source is credible enough to cite.

    Name the author. State the organization. Link to primary sources. Include dates on time-sensitive claims (“as of April 2026”). These signals tell the AI system this content has an accountable source, not anonymous text.

    6. Freshness on Time-Sensitive Topics

    For any topic where recency matters — product pricing, regulatory status, current events — AI systems heavily weight recently indexed, recently updated content. A page published April 2026 beats a page published January 2025 for queries about current status, even if the older page has higher domain authority.

    Update time-sensitive content. Add “last updated” dates. Re-publish with fresh timestamps when the underlying facts change. Freshness signals are real citation drivers for volatile topic areas.

    7. Speakable and Structured Data Markup

    Speakable schema explicitly marks the passages in your content best suited for AI extraction. It’s a direct signal to AI retrieval systems: “this paragraph is the answer.” Combined with FAQPage schema, Article schema, and HowTo schema where relevant, structured markup makes your content more parseable.

    Schema doesn’t replace the underlying structure — it reinforces it. A well-structured page with schema beats a poorly structured page with schema. But a well-structured page with schema beats a well-structured page without it.

    8. Internal Link Architecture

    AI systems that crawl the web assess topical depth partly through link structure. A page that sits within a tight cluster of related pages — all cross-linking around a topic — signals topical authority more strongly than an isolated page, even if the isolated page’s content is comparable.

    Build the cluster. The hub-and-spoke architecture is as relevant for AI citation as it is for traditional SEO. Every spoke article should link to the hub; the hub should link to every spoke.

    What Doesn’t Work

    A few patterns that are intuitively appealing but don’t translate to citation lift:

    • More content for its own sake: 5,000 words of padded content is not more citable than 900 words of dense, accurate content. AI retrieval is looking for passage quality, not page length.
    • Keyword density: Traditional keyword repetition strategies don’t make content more citable. The query match is handled at retrieval; the citation decision is about answer quality, not keyword frequency.
    • Generic authority claims: “We’re the leading experts in X” is not citable. A specific data point that demonstrates expertise is.

    The Compound Effect

    These characteristics compound. A page with a direct front-loaded answer, Q&A structure, defined terms, specific facts, clear entity signals, fresh timestamps, and schema markup sitting within a well-linked cluster is materially more citable than a page with only two or three of these characteristics. The full stack produces disproportionate results.

    For the monitoring layer: How to Track When AI Systems Cite You. For the metrics: What Is AI Citation Rate?. For the full citation monitoring guide: AI Citation Monitoring Guide.


    For the infrastructure layer: Claude Managed Agents Pricing Reference | Complete FAQ Hub.

  • AI Citation Monitoring Tools — What Exists, What Doesn’t, What We Built

    AI Citation Monitoring Tools — What Exists, What Doesn’t, What We Built

    The Lab · Tygart Media
    Experiment Nº 570 · Methodology Notes
    METHODS · OBSERVATIONS · RESULTS

    You want to monitor whether AI systems are citing your content. What tools actually exist for this, what they do, what they don’t do, and what we’ve built ourselves when nothing on the market fit.

    The Market as of April 2026

    The AI citation monitoring category is real but nascent. Here’s an honest inventory:

    Established SEO Platforms Adding AI Visibility Metrics

    Several major SEO platforms have added “AI visibility” or “AI search” modules in the past 6–12 months. These generally track:

    • Whether your domain appears in AI Overviews for tracked keywords (via SERP scraping)
    • Brand mentions in AI-generated snippets
    • Comparative visibility versus competitors in AI search results

    Ahrefs, Semrush, and Moz have all moved in this direction to varying degrees. Verify current feature availability — this has been an active development area and capabilities have changed rapidly.

    Mention Monitoring Tools Expanding to AI

    Brand mention tools like Brand24 and Mention have begun tracking AI-generated content that includes brand references. The challenge: they’re tracking brand name occurrences in crawled content, not necessarily AI citation events. Useful for brand visibility in AI-generated content that gets published, less useful for tracking in-session citations.

    Purpose-Built AI Citation Tools (Emerging)

    Several purpose-built tools targeting AI citation tracking specifically have launched or raised funding in early 2026. This category is moving fast. As of our last check:

    • Tools focused on tracking specific brand or entity mentions across AI platforms
    • API-first tools targeting developers who want to build citation monitoring into their own workflows
    • Dashboard tools with pre-built query sets for common industry categories

    Treat any specific product recommendation here as a starting point for your own research — the category will look different in 6 months.

    Google Search Console

    The strongest existing tool, and it’s free. AI Overviews that cite your pages register as impressions and clicks in GSC under the relevant queries. This is first-party data from Google itself. Limitation: covers only Google AI Overviews, not Perplexity, ChatGPT, or other platforms.

    What We Built

    When no existing tool covered the specific workflows we needed, we built our own. The stack:

    Perplexity API Query Runner

    A Cloud Run service that runs a predefined query set against Perplexity’s API on a weekly schedule. It parses the citations field from each response, checks for domain appearances, and writes results to a BigQuery table. Total engineering time: roughly one day. Ongoing cost: minimal (Cloud Run idle cost + Perplexity API usage).

    The output: a weekly BigQuery record per query showing which domains Perplexity cited, with timestamps. Trend queries show citation rate over time by query cluster.

    GSC AI Overview Monitor

    Not a custom build — just systematic review of GSC data. We check weekly which queries are generating AI Overview impressions for our tracked sites. The signal: if a page is generating AI Overview impressions on new queries, that’s a citation event.

    Manual ChatGPT Sampling

    For highest-priority queries, manual weekly sampling of ChatGPT with web search enabled. We log results to a shared spreadsheet. Less scalable than the API approach, but ChatGPT’s web search activation is inconsistent enough that API automation adds complexity without proportional reliability gain.

    What Doesn’t Exist (That Would Be Useful)

    The tool gaps that we still feel:

    • Cross-platform citation dashboard: A single view showing citation rate across Perplexity, ChatGPT, Gemini, and AI Overviews for the same query set. Nobody has built this cleanly yet.
    • Historical citation rate database: Knowing your citation rate is useful. Knowing whether it improved after you published a new piece of content is more useful. The temporal correlation is hard to establish with spot-check sampling.
    • Competitor citation tracking at scale: Easy to check manually for specific queries; hard to monitor systematically across a large competitor set and query space.

    These gaps exist because the category is new, not because the problems are technically hard. Expect the tool landscape to fill in significantly over the next 12 months.

    How to calculate citation rate: What Is AI Citation Rate?. How to set up tracking: How to Track When ChatGPT or Perplexity Cites Your Content. How to optimize for citations: How to Write Content That AI Systems Cite.


    The Perplexity API monitoring stack we built runs on Claude. For the hosted infrastructure context: Claude Managed Agents Pricing Reference | Complete FAQ.

  • What Is AI Citation Rate? (And How to Calculate Yours)

    What Is AI Citation Rate? (And How to Calculate Yours)

    AI citation rate is a metric that doesn’t have a standard definition yet, which means everyone using the term might mean something slightly different. Here’s what it is, how to calculate it, and what it actually measures — and doesn’t.

    Definition

    AI Citation Rate

    The percentage of sampled AI queries where a specific domain or URL appears as a cited source in the AI system’s response.

    Formula: (Queries where your domain appeared as a source) ÷ (Total queries sampled) × 100

    A Concrete Example

    You run 50 queries in Perplexity across your core topic cluster. Your domain appears as a cited source in 12 of those responses. Your AI citation rate for that query set on that platform: 12/50 = 24%.

    That’s the basic calculation. The complexity is in what you define as your query set, which platforms you sample, and what counts as a “citation.”

    What Counts as a Citation

    Not all AI source mentions are equal. Some distinctions worth tracking separately:

    • Direct URL citation: The AI explicitly lists your URL as a source. Highest confidence — trackable programmatically via API.
    • Domain mention: Your domain name appears in the response text but not necessarily as a formal source citation.
    • Brand mention: Your brand name appears in the response. May or may not correlate with your web content being the source.
    • Implied citation: Content clearly derived from your page but no explicit attribution. Only detectable through content fingerprinting — difficult at scale.

    For tracking purposes, direct URL citation is the most reliable signal. Brand mentions are noisier but still worth tracking for brand visibility purposes.

    How to Calculate It

    Step 1: Define Your Query Set

    Select 20–100 queries where you want to appear. Good sources for your query set:

    • Your highest-impression GSC queries (you rank for these — do AI systems cite you?)
    • Queries where you’ve published dedicated content
    • Queries from your keyword research that match your expertise
    • Questions your clients or prospects actually ask

    Step 2: Sample Across Platforms

    Run each query in Perplexity (most trackable — consistent citation format), ChatGPT with web search enabled, and Google AI Overviews (via organic search). Track results separately by platform — citation rates vary significantly between platforms for the same query set.

    Step 3: Log Results

    For each query on each platform, record:

    • Whether your domain appeared as a citation (binary: yes/no)
    • Position if ranked (first citation, third citation, etc.)
    • Date of query

    Step 4: Calculate Rate

    Aggregate by time period (weekly or monthly). Calculate separately by platform and by topic cluster — aggregate rate across all platforms and queries hides the variation that’s actually useful.

    Step 5: Establish Baseline, Then Track Change

    Your first 4–6 weeks of data sets your baseline. After that, track directional change — is the rate improving, declining, or stable? Correlate changes with content updates, new publications, and competitor activity.

    What Citation Rate Actually Measures (And Doesn’t)

    AI citation rate is a proxy for content authority signal in AI systems — not a direct ranking factor you can optimize mechanically. It reflects:

    • Whether your content is being indexed and surfaced by AI systems for your target queries
    • Whether your content structure and freshness match what AI systems prefer to cite
    • Relative authority versus competitors for the same query space

    It doesn’t measure:

    • Whether AI systems are using your content without citation (training data influence)
    • User behavior after AI responses (do they click through to your site?)
    • Revenue impact of being cited (cited ≠ converting)

    Benchmarks and Context

    Because this metric is new, industry benchmarks don’t exist yet. What matters is your own trend line, not comparison to a published standard. A 20% citation rate in a highly competitive topic cluster might represent strong performance; 20% in a niche you should dominate might indicate underperformance. Context is everything.

    For the full monitoring setup: How to Track When ChatGPT or Perplexity Cites Your Content. For tools available: AI Citation Monitoring Tools Comparison. For content optimization: How to Write Content That AI Systems Actually Cite.


    For the agent infrastructure behind automated citation tracking: Claude Managed Agents Pricing and FAQ Hub.

  • Types of Radon Mitigation Systems Explained

    Types of Radon Mitigation Systems Explained

    The Distillery — Brew № 1 · Radon Mitigation

    There is no single radon mitigation system. There are six primary system types, each designed for specific foundation conditions — and most homes with elevated radon require one primary method plus supplemental sealing. Knowing which system type applies to your home’s foundation eliminates confusion about what a contractor is proposing and whether the approach matches your situation.

    1. Active Sub-Slab Depressurization (ASD)

    Active Sub-Slab Depressurization is the most widely installed radon mitigation system in the United States. It is the standard approach for slab-on-grade homes and basement homes with concrete slab floors.

    How ASD Works

    A suction pipe penetrates the concrete slab, connecting to the aggregate or soil layer beneath. A continuously running electric fan draws air (and with it, radon) from beneath the slab, routing it through PVC pipe to discharge above the roofline. This creates negative pressure in the sub-slab zone relative to the home’s interior — preventing radon from finding pathways through cracks, joints, and penetrations into the living space.

    ASD Applications

    • Slab-on-grade homes (full footprint slab, no basement)
    • Basement homes with concrete slab floors
    • Homes with both a basement and upper-level slab additions
    • Garage slabs connected to the main living area slab

    ASD Governing Standard

    AARST-ANSI SGM-SF (Standard of Practice for Mitigation of Radon in Schools and Large Buildings, adapted for single-family) governs ASD installation requirements including diagnostic testing, pipe sizing, fan placement, and performance verification.

    2. Active Sub-Membrane Depressurization (ASMD)

    Active Sub-Membrane Depressurization is the crawl space equivalent of ASD. Instead of drilling through concrete, the system creates negative pressure beneath a vapor barrier (membrane) installed over the crawl space soil.

    How ASMD Works

    A heavy-duty polyethylene vapor barrier (minimum 6-mil; professional installations use 10–20 mil) is installed across the entire crawl space floor, lapped up foundation walls, and sealed at all edges and penetrations. A suction pipe penetrates the barrier and connects to the soil or aggregate below via a perforated collection mat. The fan draws soil gas from beneath the barrier, routing it above the roofline through the same type of PVC pipe system used in ASD.

    ASMD Requirements

    • Foundation vents must be sealed — open vents allow outdoor air into the crawl space, defeating the sub-membrane vacuum
    • Barrier seams must be lapped (minimum 12″ overlap) and taped
    • Multiple suction points are often needed — crawl spaces typically require 2–4 collection points versus the 1–2 typical in ASD installations
    • AARST-ANSI RMS-LB governs ASMD installation standards

    3. Drain-Tile Depressurization

    Many basement homes — particularly those built after 1980 — were constructed with a drain-tile system: a perforated pipe network running around the interior or exterior perimeter of the foundation, at or below the footing level, designed to channel groundwater to a sump pit. This drain tile can serve as a highly effective radon collection network.

    How Drain-Tile Depressurization Works

    When a sump pit is present and the drain tile is functional, the mitigator creates suction at the sump pit — either by sealing the pit with an airtight lid and connecting a fan, or by installing a dedicated suction pipe into the drain tile network. Because the drain tile runs around the full foundation perimeter, a single suction point at the sump can create negative pressure across a very large area — often the entire foundation footprint without any slab drilling.

    Advantages Over Standard ASD

    • No slab drilling required (the drain tile network is already in place)
    • Often achieves better sub-foundation coverage than a single slab core hole
    • Sump pit is already present — lid modification is the primary work
    • Lower installation cost when drain tile is accessible

    Limitations

    • Requires a confirmed functional drain-tile system — older or poorly maintained tile may be silted or blocked
    • Not present in all homes — many older homes and slab-on-grade construction have no drain tile
    • May need to be supplemented with slab suction point(s) if tile coverage is incomplete

    4. Block-Wall Depressurization

    Concrete masonry unit (CMU) block foundation walls have hollow cores that communicate directly with the soil — a significant secondary radon entry pathway in older homes. Block-wall depressurization addresses this specifically.

    How Block-Wall Depressurization Works

    Small holes (2″–3″ diameter) are drilled through the interior face of the CMU block wall, typically just above the slab level, at 6–8 foot intervals around the affected perimeter. PVC pipe connects these holes, manifolding into the main ASD fan system or a dedicated fan. The fan draws radon from inside the block core cavities before it can migrate through mortar joints and wall cracks into the basement air.

    When Block-Wall Depressurization Is Needed

    • Post-mitigation testing still shows levels above 4.0 pCi/L after standard ASD is installed
    • Visual inspection reveals significant efflorescence, spalling, or moisture infiltration through block walls (indicating active soil gas pathways)
    • Home is pre-1975 CMU construction with no poured concrete wall facing

    Block-wall depressurization is almost always an add-on to ASD, not a standalone system. Cost: $300–$600 in additional materials and labor when added to an existing ASD installation.

    5. Heat Recovery Ventilator (HRV) or Energy Recovery Ventilator (ERV)

    HRV and ERV systems are whole-house mechanical ventilation systems that exchange stale indoor air with fresh outdoor air while recovering heat (HRV) or both heat and moisture (ERV). They are sometimes used as a radon reduction strategy — primarily in situations where other methods are impractical or as a supplemental approach.

    How HRV/ERV Reduces Radon

    By continuously introducing fresh outdoor air into the home, HRV/ERV dilutes indoor radon concentrations. They also reduce the negative pressure differential that draws radon into the home from the soil, because they balance indoor and outdoor pressure rather than allowing the home to depressurize relative to the soil.

    Limitations as Radon Mitigation

    • Less reliable reduction than ASD/ASMD — radon dilution depends on outdoor air exchange rate, and results vary significantly by climate and home tightness
    • Higher operating cost — HRV/ERV units consume 100–400 watts versus 20–90 watts for a radon fan
    • Does not address the root cause (radon entry from soil) — only dilutes after entry
    • Not accepted as primary mitigation in all state radon programs
    • Best suited as supplemental to ASD in homes where additional air quality improvement is also desired

    EPA and AARST consider ASD/ASMD the preferred primary mitigation method. HRV/ERV may be appropriate as supplemental mitigation or in unusual foundation situations where ASD is genuinely impractical.

    6. Natural Ventilation Enhancement

    Natural ventilation — opening windows, operating exhaust fans, increasing air exchange — can temporarily reduce radon concentrations. It is not a mitigation system and is not recommended by EPA or AARST as a radon control strategy for several reasons:

    • Effective only while windows are open — unpractical in most U.S. climates for the majority of the year
    • Increases heating and cooling costs significantly
    • Can create negative pressure that worsens radon entry
    • Provides no permanent solution

    Natural ventilation may be used as a short-term measure while a permanent system is being installed, but it is not a substitute for ASD, ASMD, or other mechanical systems.

    Choosing the Right System: Decision Guide

    Foundation TypePrimary SystemCommon Add-On
    Slab-on-gradeASDSealing (cracks, joints)
    Basement — poured concreteASDDrain-tile depressurization if sump present
    Basement — CMU block wallsASDBlock-wall depressurization
    Crawl space — ventedASMD (with encapsulation)Foundation vent sealing
    Crawl space — encapsulatedASMDAdditional suction points if needed
    New construction (RRNC)Passive pipe (fan-ready)Fan activation if post-construction test elevated
    Combination foundationASD + ASMD (separate systems or manifolded)Sealing at transition zones

    Frequently Asked Questions

    What is the most common type of radon mitigation system?

    Active Sub-Slab Depressurization (ASD) is the most commonly installed radon mitigation system in the U.S. It applies to slab-on-grade and basement homes — the two most prevalent residential foundation types. For crawl space homes, Active Sub-Membrane Depressurization (ASMD) is the standard.

    Can one system work for multiple foundation types in the same home?

    Yes, but it typically requires separate or manifolded systems. A home with a basement and a slab-on-grade addition, for example, may need ASD suction points in both zones, connected to a single fan via manifold pipe — or two separate fans if the zones are not contiguous. An experienced mitigator will design for the full footprint, not just the primary foundation type.

    Does the type of radon system affect the cost?

    Yes, significantly. A standard single-point ASD in a poured concrete basement is the least expensive ($800–$1,500). Adding drain-tile depressurization at the sump typically adds $100–$300. Block-wall depressurization adds $300–$600. ASMD with full crawl space encapsulation can run $2,500–$5,000+ depending on crawl space size and membrane quality.

    What type of radon system works in a home with no basement and no crawl space?

    Slab-on-grade homes use ASD — a suction pipe drilled through the concrete slab connects to the aggregate beneath. Interior routing typically runs through a garage wall or utility closet to the attic. Exterior routing is an alternative when interior access is limited. The challenge in slab homes is pipe routing to above the roofline without a basement or crawl space to work through — but it is fully achievable in almost all cases.

    What is the difference between ASD and ASMD?

    Both use a fan to create negative pressure below the home’s floor system. ASD drills through a concrete slab and draws suction from the sub-slab aggregate or soil. ASMD installs a vapor barrier over the crawl space soil and draws suction from beneath the barrier — no concrete is present to drill through. The fan, pipe, and discharge components are identical; only the suction connection method differs.

  • How to Track When ChatGPT or Perplexity Cites Your Content

    How to Track When ChatGPT or Perplexity Cites Your Content

    Tygart Media Strategy
    Volume Ⅰ · Issue 04Quarterly Position
    By Will Tygart
    • Long-form Position
    • Practitioner-grade

    ChatGPT cited a competitor’s blog post instead of yours. Perplexity summarized the wrong article. An AI answer engine described your service category without mentioning you. You’d like to know when this happens — and whether it’s improving over time.

    The problem: no one has built a clean, turnkey tool for this yet. Here’s what actually exists, what we’ve pieced together, and what a real tracking setup looks like.

    Why This Is Hard

    Web search citation tracking is solved: rank trackers like Ahrefs and SEMrush show you who’s linking to what. AI citation tracking has no equivalent infrastructure. Here’s why:

    • Non-deterministic outputs: Ask ChatGPT the same question twice; you may get different sources cited, or no sources at all. There’s no persistent ranking to track.
    • No public citation index: Google’s index is crawlable. There’s no equivalent for “content that AI systems have cited in responses.” You can’t pull a report.
    • Variable source disclosure: Perplexity shows sources. ChatGPT’s web-enabled mode shows sources sometimes. Gemini shows sources. Claude generally doesn’t show sources in the same way. Tracking works where sources are disclosed; it breaks where they aren’t.
    • Query sensitivity: Your content might get cited for one phrasing and completely missed for a near-synonym. There’s no search volume data to tell you which phrasings matter.

    What Actually Exists Today

    Manual Query Sampling

    The only fully reliable method: run queries yourself and check the sources cited. For a content monitoring program this might look like:

    • Define 20–50 queries where you want to appear (covering your core topics)
    • Run each query in Perplexity, ChatGPT (web-enabled), and Gemini weekly or biweekly
    • Log whether your domain appears in cited sources
    • Track citation rate (appearances / total queries run) over time

    This is tedious but gives you ground truth. It’s what a real monitoring program looks like before you automate it.

    Perplexity Source Tracking

    Perplexity consistently displays its sources, making it the most tractable platform for systematic citation tracking. A simple automated approach:

    • Use Perplexity’s API to query your target questions programmatically
    • Parse the citations field in the response
    • Check whether your domain appears
    • Log and aggregate over time

    Perplexity’s API is available with a subscription. The citations field returns the URLs Perplexity used to generate its answer. You can run this as a scheduled Cloud Run job and dump results to BigQuery for trend analysis.

    ChatGPT Web Search Mode

    When ChatGPT uses web search (either via the browsing tool or search-enabled API), it returns source citations. The search-enabled ChatGPT API (available with OpenAI API access) gives you programmatic access to these citations. Same approach: define queries, run them, parse citations, track your domain.

    Limitation: not all ChatGPT responses use web search. For queries it answers from training data, no source is cited and you have no visibility into whether your content influenced the answer.

    Google AI Overviews

    Google AI Overviews (formerly SGE) shows cited sources inline in search results. You can track these through Google Search Console for your own content — if Google’s AI Overview cites your page, that page gets an impression and potentially a click recorded in GSC under that query. This is the only AI citation signal with first-party tracking infrastructure.

    Emerging Tools

    As of April 2026, several tools are building toward AI citation tracking as a category: mention monitoring services that have added AI search coverage, SEO platforms adding “AI visibility” metrics, and purpose-built tools targeting this specific problem. The category is forming but not mature. Verify current capabilities — this space has changed significantly in the past six months.

    What a Real Monitoring Setup Looks Like

    Here’s the practical stack we’ve assembled for tracking citation presence across AI platforms:

    1. Define your query set: 30–50 queries across your core topic clusters. Weight toward queries where you have existing content and where you’re trying to establish authority.
    2. Perplexity API integration: Scheduled weekly run. Parse citations. Log domain appearances to a tracking spreadsheet or BigQuery table.
    3. ChatGPT web search sampling: Less systematic — manual sampling weekly for highest-priority queries. The API approach works but requires more engineering to handle variability in when web search activates.
    4. Google Search Console: Monitor AI Overview impressions. This is your strongest signal because it’s Google’s own data, not sampled queries.
    5. Baseline and trend: After 4–6 weeks of tracking, you have a baseline citation rate. Changes correlate (imperfectly) with content quality improvements, new publications, and competitor activity.

    What Citation Rate Actually Tells You

    Citation rate — your domain appearances divided by total queries sampled — is a proxy metric, not a direct ranking signal. What drives it:

    • Content freshness: AI systems prefer recently indexed, recently updated content for queries about current information
    • Structural clarity: Content with explicit Q&A structure, defined terms, and direct factual claims gets cited more reliably than narrative content
    • Domain authority signals: The same signals that help SEO rankings help AI citation rates — but the weighting may differ by platform
    • Entity specificity: Content that clearly establishes your brand as an entity with defined characteristics gets cited more consistently than generic content

    For the content optimization angle: AI Citation Monitoring Guide. For the broader GEO picture: What Managed Agents means for content visibility.

    For the hosted agent infrastructure context: Claude Managed Agents Pricing Reference — how the billing works for agents that could automate citation monitoring workflows.

  • The Real Monthly Cost of Running Claude Managed Agents 24/7

    The Real Monthly Cost of Running Claude Managed Agents 24/7

    Tygart Media Strategy
    Volume Ⅰ · Issue 04Quarterly Position
    By Will Tygart
    • Long-form Position
    • Practitioner-grade

    If you’re considering running Claude Managed Agents around the clock, you want a number. Not “it depends.” An actual number you can put in a budget. Here’s the math, worked out by scenario, with the honest caveats about where the real costs are.

    The Formula

    Total monthly cost = (Active session hours × $0.08) + token costs + optional tool costs

    The $0.08/session-hour charge only applies during active execution. Idle time — waiting for input, tool confirmations, external API responses — doesn’t count. This matters significantly for 24/7 workloads, because very few agents are active 100% of the time even when “running around the clock.”

    The Maximum Theoretical Cost

    Scenario: Agent running continuously, zero idle time, 24 hours a day, 30 days a month.

    • Session runtime: 24 hrs × $0.08 × 30 days = $57.60/month
    • Token costs: separate, highly variable (see below)

    $57.60/month is the ceiling on session runtime charges. You cannot pay more than this in session fees under any 24/7 scenario. But here’s the reality: that ceiling assumes zero idle time across the entire month, which doesn’t describe any real production agent.

    Realistic 24/7 Scenarios

    Monitoring Agent (High Idle Ratio)

    Runs continuously watching for triggers — error alerts, specific data patterns, incoming requests. Activates on trigger, processes, returns to monitoring state.

    • Assumption: 5% active execution time (watching 95% of the time, executing 5%)
    • Active hours: 24 × 30 × 0.05 = 36 hours/month
    • Session runtime: 36 × $0.08 = $2.88/month
    • Token costs: low — moderate bursts on trigger events
    • Realistic total: $5–15/month

    Customer Support Agent (Business Hours Active)

    “24/7” in the sense of always-available, but actual request volume concentrates in business hours. Waits for tickets, processes them, waits again.

    • Assumption: 8 hours/day active execution, 16 hours waiting
    • Active hours: 8 × 30 = 240 hours/month
    • Session runtime: 240 × $0.08 = $19.20/month
    • Token costs: depends heavily on ticket volume and average length
    • At 100 tickets/day with moderate length: likely $30–80/month in tokens
    • Realistic total: $50–100/month

    Continuous Autonomous Pipeline

    Batch processing agent that runs continuously through a queue with minimal waiting — the closest to true 24/7 active execution.

    • Assumption: 20 hours/day truly active (4 hours queue exhaustion/maintenance)
    • Active hours: 20 × 30 = 600 hours/month
    • Session runtime: 600 × $0.08 = $48/month
    • Token costs: high — continuous processing means continuous token consumption
    • This is where tokens become the dominant cost driver by a significant margin
    • Realistic total: $200–500+/month (tokens dominate)

    The Real Variable: Token Costs

    For any 24/7 workload that’s genuinely busy, token costs will substantially exceed session runtime costs. The math:

    A moderately active agent processing 10,000 input tokens and 2,000 output tokens per hour with Claude Sonnet 4.6:

    • Input: 10,000 tokens × $3/million = $0.03/hour
    • Output: 2,000 tokens × $15/million = $0.03/hour
    • Token cost: $0.06/hour vs. session runtime of $0.08/hour — roughly equal at this volume

    Scale to 100,000 input tokens and 20,000 output tokens per hour (a busy processing agent):

    • Input: $0.30/hour; Output: $0.30/hour
    • Token cost: $0.60/hour vs. session runtime of $0.08/hour — tokens are 7.5× the runtime charge

    The session runtime fee is flat and bounded. Token costs scale with workload volume. For high-volume 24/7 agents, optimize token efficiency (prompt caching, context management, output brevity) before worrying about the session runtime charge.

    Prompt Caching Changes the Token Math

    If your agent has a large, stable system prompt — common in agents with extensive tool definitions or knowledge bases — prompt caching dramatically reduces input token costs. Cache hits cost a fraction of base input rates. For a 24/7 agent with a 20,000-token system prompt hitting the same context repeatedly, caching that prompt can cut input costs by 80–90%. The session runtime charge is unchanged, but the total cost picture improves significantly.

    The Budget Summary

    Agent Type Runtime/mo Typical Total
    Monitoring / low activity ~$3 $5–15
    Support agent (business hours volume) ~$19 $50–100
    Continuous processing pipeline ~$48 $200–500+
    Theoretical maximum (zero idle) $57.60 Unbounded (tokens)

    Complete pricing reference: Claude Managed Agents Pricing Guide. How idle time affects billing: Idle Time and Billing Explained. All questions: FAQ Hub.

    What to do next

    Now that you have the cost math — here’s how to choose and implement

    You now know what Managed Agents costs at scale. The next decision is whether it’s the right architecture vs. OpenAI’s equivalent — and what the implementation actually looks like in practice.

  • Claude Managed Agents vs. OpenAI Agents API — A Direct Comparison

    Claude Managed Agents vs. OpenAI Agents API — A Direct Comparison

    TL;DR — Pick one in 30 seconds

    Choose Claude Managed Agents for zero-infra, fast production deployment. Choose OpenAI Agents API if you need multi-model flexibility or already run on OpenAI infrastructure.

    Feature Claude Managed Agents OpenAI Agents API
    Model lock-in Claude only GPT-4o, o3 — OAI only
    Setup complexity Zero infra — fully managed SDK — you build the harness
    Memory Built-in (public beta, May 2026) Manual via vector DB
    Multiagent Native (lead + specialists) Swarm/SDK patterns
    Pricing $0.08/session-hr + tokens Token-only (no session fee)
    Best for Fast production, Claude-native Multi-model, existing OAI infra

    Model Accuracy Note — Updated May 2026

    Current flagship: Claude Opus 4.7 (claude-opus-4-7). Current models: Opus 4.7 · Sonnet 4.6 · Haiku 4.5. Claude Opus 4.6 referenced in this article has been superseded. See current model tracker →

    Tygart Media Strategy
    Volume Ⅰ · Issue 04Quarterly Position
    By Will Tygart
    • Long-form Position
    • Practitioner-grade

    You’re evaluating hosted agent infrastructure. Both Anthropic and OpenAI have one. Before you commit to either, here’s what’s actually different — not the marketing version, the architectural and pricing version.

    Bottom Line Up Front

    If your stack is Claude-native and you want to get to production fast without building orchestration infrastructure, Managed Agents is hard to beat. If you need multi-model flexibility or have OpenAI deeply embedded in your stack, the calculus changes. Lock-in is real on both sides.

    Still Deciding?

    I’ve run both. Email me your use case and I’ll tell you which one fits.

    No pitch. If Claude isn’t the right call for what you’re building, I’ll tell you that too.

    Email Will → will@tygartmedia.com

    What Each Product Is

    Claude Managed Agents

    Anthropic’s hosted runtime for long-running Claude agent work. You define an agent (model, system prompt, tools, guardrails), configure a cloud environment, and launch sessions. Anthropic handles sandboxing, state management, checkpointing, tool orchestration, and error recovery. Launched April 8, 2026 in public beta.

    OpenAI Agents API

    OpenAI’s hosted agent infrastructure layer, launched earlier in 2026. Provides similar capabilities: hosted execution, tool integration, multi-agent coordination. Supports multiple OpenAI models (GPT-4o, o1, o3, etc.).

    Model Flexibility

    Managed Agents: Claude models only. Sonnet 4.6 and Opus 4.6 are the primary options for agent work. No multi-model mixing within the managed infrastructure.

    OpenAI Agents API: OpenAI models only, but a wider current model lineup (GPT-4o, o1, o3-mini depending on task). Also Claude-only within its own ecosystem — not multi-model in the cross-provider sense.

    The practical implication: If your evaluation is “I want the best model for this specific task regardless of provider,” neither hosted solution gives you that. Both lock you to their provider’s models. The multi-model comparison matters for self-hosted frameworks (LangChain, etc.), not for managed hosted solutions.

    Pricing Structure

    Claude Managed Agents: Standard Claude token rates + $0.08/session-hour of active runtime. Idle time doesn’t bill. Code execution containers included in session runtime — not separately billed.

    OpenAI Agents API: Standard OpenAI token rates + usage-based tooling costs. Pricing structure varies by tool and model tier. Verify current rates at OpenAI’s pricing page — rates have changed multiple times as their agent products have evolved.

    Direct comparison difficulty: Without modeling the same specific workload against both providers’ current rates, headline comparisons mislead. Token rates differ by model, model capabilities differ, and “session runtime” isn’t a category OpenAI uses. Model the workload, not the headline number.

    Infrastructure and Lock-In

    Both solutions create meaningful lock-in. This isn’t a criticism — it’s an honest description of the trade-off you’re making:

    Claude Managed Agents lock-in: Your agents run on Anthropic’s infrastructure with their tools, session format, sandboxing model, and checkpointing. Migrating to OpenAI’s Agents API or self-hosted infrastructure requires rearchitecting session management, tool integrations, and guardrail logic. One developer’s reaction at launch: “Once your agents run on their infra, switching cost goes through the roof.”

    OpenAI Agents API lock-in: Symmetric. Same dynamic in reverse. OpenAI’s session format, tool integration patterns, and infrastructure assumptions create equivalent switching costs to move to Anthropic’s platform.

    The honest framing: You’re not choosing “open” vs. “locked.” You’re choosing which provider’s lock-in you’re more comfortable with, given your existing infrastructure, model preferences, and vendor relationship.

    Data Sovereignty

    Both solutions run your data on provider-managed infrastructure. Neither currently offers native on-premise or multi-cloud deployment for the managed hosted layer. For companies with strict data sovereignty requirements, this is a parallel constraint on both platforms — not a differentiator.

    Production Track Record

    Claude Managed Agents: Launched April 8, 2026. Production users at launch: Notion, Asana, Rakuten (5 agents in one week), Sentry, Vibecode, Allianz. Anthropic’s agent developer segment run-rate exceeds $2.5 billion.

    OpenAI Agents API: Earlier launch gives more time in production, but the product has been revised significantly since initial release. Longer production history, but also more legacy architectural assumptions baked in.

    When to Choose Claude Managed Agents

    • Your stack is already Claude-native (you’re using Sonnet or Opus for most model calls)
    • You want to reach production without building orchestration infrastructure
    • Your tasks are long-running and asynchronous — the session-hour model fits naturally
    • The Notion, Asana, or Sentry integrations are relevant to your workflow
    • You want Anthropic’s specific safety and reliability guarantees

    When to Consider OpenAI’s Agents API Instead

    • Your stack is already heavily OpenAI-integrated (GPT-4o for primary model work, existing tool integrations)
    • You need access to reasoning models (o1, o3) for specific task types — Anthropic’s equivalent is Claude’s extended thinking, which has different characteristics
    • The specific tool integrations in OpenAI’s ecosystem are better matched to your stack
    • You want more production time at scale before committing to a platform

    When to Use Neither (Self-Hosted Frameworks)

    LangChain, LlamaIndex, and similar self-hosted frameworks remain viable — and better — when you genuinely need multi-model flexibility, on-premise execution, or tighter loop control than either hosted solution provides. The trade-off is engineering effort: months of infrastructure work that Managed Agents or OpenAI’s API eliminates.

    Complete pricing breakdown: Claude Managed Agents Pricing Reference. All Managed Agents questions: FAQ Hub. Enterprise deployment example: Rakuten: 5 Agents in One Week.

  • How Claude Managed Agents Handles Idle Time (And Why It Matters for Your Bill)

    How Claude Managed Agents Handles Idle Time (And Why It Matters for Your Bill)

    Tygart Media Strategy
    Volume Ⅰ · Issue 04Quarterly Position
    By Will Tygart
    • Long-form Position
    • Practitioner-grade

    The most counterintuitive thing about Claude Managed Agents pricing is what you don’t pay for. Most people, when they hear “$0.08 per session-hour,” mentally model a virtual machine running continuously. That’s the wrong mental model. Here’s the right one, and why it matters for your bill.

    The Core Distinction: Active vs. Idle

    Managed Agents session runtime only accrues while your session’s status is running. The session can exist — open, initialized, capable of continuing — without accumulating runtime charges when it’s not actively executing.

    The specific states that do not count toward your $0.08/hr charge:

    • Time spent waiting for your next message
    • Time waiting for a tool confirmation
    • Time waiting on an external API response your tool is calling
    • Rescheduling delays
    • Terminated session time

    This is a meaningful architectural decision by Anthropic. They’re billing on what actually taxes their compute — active execution — not on session existence or wall-clock time.

    Why This Is Different From How You Might Expect Billing to Work

    Compare three billing models:

    Virtual machine billing (what this is not): You pay for every hour the instance exists, whether it’s idle or saturated. A VM running 24/7 with 10% actual utilization still costs 24 hours/day.

    Lambda/function billing (closer analogy): AWS Lambda bills on execution duration and invocation count — you pay when code actually runs, not when a function is “available.” Idle Lambda functions cost nothing.

    Managed Agents billing (what this actually is): Closer to Lambda than VM. You pay $0.08 per hour of active execution. A session that runs for 2 hours of wall-clock time but has 90 minutes of waiting costs $0.08 × 1.5 hours = $0.12, not $0.08 × 2 hours = $0.16.

    A Real Scenario: The Human-in-the-Loop Agent

    Consider an agent that processes your inbox for action items and waits for your approval before sending replies. Wall-clock time: 4 hours open during your workday. Actual active execution: 20 minutes of processing across that 4-hour window, with the rest spent waiting for your review decisions.

    • VM billing equivalent: 4 hours × rate = significant charge
    • Managed Agents billing: 20 minutes × $0.08/hr = $0.027

    The difference is real. For interaction-heavy agents where the agent frequently waits for human decisions, the idle-time exclusion significantly reduces costs versus a naive per-hour model.

    A Real Scenario: The Autonomous Batch Agent

    Now consider an agent running a fully autonomous content pipeline — no human checkpoints, just continuous execution through a queue. Wall-clock time and active execution time are nearly identical because the agent never waits.

    • A 2-hour autonomous batch: 2 hours × $0.08 = $0.16

    Here, the idle-time model provides no benefit — the agent has no idle time. The billing is effectively equivalent to per-hour pricing because execution is continuous.

    Code Execution Containers Are Included

    One more billing nuance worth knowing: when your agent runs code, the execution happens in sandboxed Linux containers. These containers are not separately billed on top of session runtime. The $0.08/hr covers both the session runtime and the container execution. This is explicitly documented by Anthropic and represents meaningful savings if your agent is doing significant code execution work — you’re not paying twice.

    What This Means for Workload Design

    If you’re designing agent workflows and have the choice between architectures, the billing model creates a useful signal:

    • Agents that wait on humans: Metered billing is favorable — you only pay for the actual reasoning and execution time, not the human decision time
    • Fully autonomous agents: Billing approaches equivalent to per-hour rates — optimize these on token efficiency, not idle reduction
    • Scheduled batch agents: Natural fit — run when needed, terminate when done, no idle accumulation

    The 24/7 Agent Math

    For anyone doing the 24/7 always-on calculation: the maximum theoretical runtime exposure is 24 hrs × $0.08 × 30 days = $57.60/month in session fees. But a 24/7 agent with zero idle time is rare in practice. Agents that sleep between triggers, wait on external data, or hold for human decisions have meaningful idle windows that reduce the actual charge below the theoretical ceiling.

    Full monthly cost analysis: The Real Monthly Cost of Running Claude Managed Agents 24/7. Pricing reference: Complete Pricing Guide. All questions: FAQ Hub.