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

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.