Tag: AI Agents

  • Review the Run, Not the Diff: Cursor’s Shareable Canvases in a Fleet Loop

    Review the Run, Not the Diff: Cursor’s Shareable Canvases in a Fleet Loop

    What shipped on October 16

    Cursor’s changelog entry dated October 16, 2026 — “Cloud Agent Canvases and Environment Improvements” — is three short sections long, and one of them changes the shape of a finished run: “Cloud agents can now build interactive canvases as part of their work. Share a finished canvas with your team by sending its link.” That sentence is Cursor’s. Everything in this section is theirs, from that changelog entry and the Canvases page of their docs, read October 19, 2026.

    A canvas, in their docs, is an interactive artifact that renders next to the chat: a standalone view laid out with sections, stats, and tables, instead of a long markdown table or a code block to scroll through. Ask an agent for a dashboard, an analysis, an audit, or a report, and Cursor opens the result in a canvas when a visual view is the better fit. The canvas is saved to the workspace’s canvas list, so you can reopen it, edit it, and rerun it later with fresh data.

    Sharing is a publish, not a forward. From the canvas toolbar you Publish — which publishes a share or refreshes an existing one — then copy the link and send it. What the link opens is a live snapshot: the same layout, charts, and tables, without rerunning the agent and without digging through the chat history. Teammates open it in the browser with read-only access. A Shared Canvases page on the dashboard lists the shares you published yourself, so you can reopen, copy, or delete your own; it does not list a teammate’s shares — for those, you ask for the link.

    The same release carried two more items, logged here for completeness. The Environments API is available in beta: a programmatic way to create and manage the cloud environments agents run in. And the environments themselves grew: up to 300 repositories per environment, with large environments loading 5x faster — Cursor’s numbers, not measurements of ours. Neither item is this piece’s subject; both are on the desk’s check list.

    The gate is part of the story

    A canvas link is not a public link. The docs are explicit: anyone on your team can open the link in the browser. Only members of your team can open a share, and you need to be on a team to share at all. Sending the link outside the team sends a door the recipient cannot walk through — which makes the canvas a team surface, not a publishing surface.

    Sharing is also a paid-plan surface with conditions attached. Shares can be created on paid plans — Premium and higher, Teams, Enterprise, and the legacy Pro, Pro+, and Ultra tiers; free accounts cannot create shares. Sharing requires a privacy mode that allows data storage, so Legacy Privacy Mode blocks it. And a team admin can turn shared canvases off for the whole organization from team settings. A shop adopting the feature checks those three gates before it builds a habit on them.

    A snapshot is not the run itself. The canvas shows the view as published; the chat, the reasoning, and the diff stay behind it. When the underlying numbers move, the share does not move with them — refreshing it is another Publish. A reviewer who needs the execution record still goes to the record. The canvas is the board on the wall, not the flight recorder.

    Where it lands in our loop

    Our loop fits in one sentence: a work order is written with everything the executor needs, the agent flies it, and the result is debriefed at the receiving end — the live page checked in a browser, the delivered file opened, the count re-added by hand. The debrief is where quality is decided here, and its surface depends on the seat. A cloud run launched over the API hands back result text. A Desktop sitting hands back a transcript and screenshots at verification points, which this shop trusts over any summary.

    A canvas inserts a third surface into that debrief: the agent arranges its own evidence — sections, stats, tables — into the board the reviewer actually opens. For the data-heavy work this shop runs constantly — citation pulls, model-catalog diffs, site audits — a rendered board is the shape the answer wanted anyway. Cursor’s docs make the same point from their side: canvases earn their keep on data-intensive tasks, where information arranged non-linearly is easier to digest than a wall of text.

    The approval moment in this shop is a tap on finished work. Will reviews results, not work in progress, and the decision lands on an artifact he can read in one pass. A canvas link is the first Cursor surface shaped exactly like that decision: open the link, read the board, decide. No repo, no diff, and no IDE sits between the finished result and the person approving it.

    The lane question the feature raises is a seat question. What reaches Stefani and Ziggy today is finished artifacts — live pages, PDFs, one-line summaries — because they work outside the operator’s seat. A canvas would be the cheapest of those artifacts to produce and the easiest to read, if the recipients hold team seats. The docs make team membership the door, so who holds a seat becomes a setup decision to make on purpose — not an assumption to discover after a link bounces off someone it was meant for.

    What is not established, labeled plainly: our production lane launches cloud agents over the API and receives text back. The docs describe canvases in the app surfaces — the Agents Window, the canvas toolbar, the dashboard. Whether a run launched through the API can produce a canvas, or return a share link, is not in the docs we read and not in our records. That question goes on the check list. It does not go into the workflow.

    One more mapping, labeled as intent rather than record. The docs say common canvas workflows can be packaged as skills, so every run that uses the skill returns the same output shape. In this shop’s language, that is a debrief template: the same board for every fleet run — what was asked, what landed, what the counts say, what failed. A debrief-canvas skill is the first experiment this change earns here. It has not run yet.

    Our record, stated plainly

    No canvas has appeared in our runs yet. Our run record — every cloud agent this desk has launched, reconciled from the API and watched daily — contains no canvas artifact and no share link as of October 19, 2026. The only “canvas” in that record is an agent named for a Google Maps canvas, a different sense of the word, and a standing reminder to grep carefully before claiming a first.

    So this piece is a map, written before the landing: what the feature is on the vendor’s own pages, what its gates are, and where it would sit in a loop that already exists. The check is decided. The next data-heavy Desktop sitting asks for its debrief as a canvas, by name, and the canvas gets read against the same sitting’s transcript — the surface we trust — before the desk adopts it anywhere. If the board and the record disagree, the record wins, and the disagreement gets written down on this page with a date.

    Method & settings — this piece

    Sources: Cursor changelog, entry dated October 16, 2026; Cursor Docs, “Canvases” (read October 19, 2026). Vendor facts are stated as the vendor’s; the secondary write-ups that surfaced this item were used as leads only, never as sources

    Our surfaces in this mapping: Desktop 3.25.9 on the laptop (read from the install itself, October 19, 2026); cloud agents over the API (model catalog baseline 46 models, October 16, 2026)

    Our canvas record: none — no canvas artifact or share link in our run record as of October 19, 2026. This is a mapping piece, not a hands-on report

    Desk rule: every fact carries its provenance and date — our record, or Cursor’s page, never blended

    Filed under the Cursor desk — the desk that runs Cursor as a production harness.

  • What’s New in Cursor Is a Claim. Here’s How We Check It.

    What’s New in Cursor Is a Claim. Here’s How We Check It.

    A changelog entry is a claim

    When a platform ships an upgrade, the changelog tells you what the vendor wants you to believe about it. Sometimes the claim is the whole story. Sometimes the feature is real and irrelevant to your work. Sometimes — the dangerous case — something you depend on moved and the entry is three lines long. A shop that builds on a harness needs a discipline that sorts those three piles. This is ours.

    The loop, in order

    • Watch. Two daily diffs: the model catalog (running since October 16, 2026) and the changelog itself (from today). Change-only reporting — silence means nothing moved, and a failed watch reports itself once rather than going quietly blind.
    • Read against our record. Every real change gets one question: does it touch a surface we run — Desktop sittings, the cloud API, mobile steering, or a model our routing names? If not, it is logged and skipped, however shiny.
    • Check before adopting. A change that touches the harness earns a check against a known quantity — a brief we have already graded, a sitting pattern we already run. The check’s result, not the changelog’s adjectives, decides what changes in our work-order kit.
    • Write it down, dated. The desk pages update with the date of the change and the date of our check. A claim we repeated and later found wrong gets corrected in place, with the correction visible.

    This week’s page, sorted by the discipline

    Cursor’s changelog top, as of October 19, 2026 (entry dates October 16 and 13 at the head of the page): cloud-agent entries lead — canvases from cloud agents, an environments API. Those matter here, because cloud agents are a production lane for us; they go on the check list. A remote-control entry — see and reply to local agents from the iOS app, in Cursor’s description — matters specifically to a shop whose sittings run on one laptop: it changes how a sitting can be supervised, so it gets read carefully before it gets used. A marketplace launch and a pair of shipping bots are logged-and-skipped for now: real, announced, and not touching our record yet. Those descriptions are Cursor’s; the sorting is ours.

    Say where each fact came from

    One habit does more for a desk’s credibility than any other: provenance on the page. The changelog currently shows entry dates but no Desktop version string — so this desk’s version knowledge does not come from the changelog. It comes from reading our own install (3.25.9, October 19, 2026). When the basis of a number is our fleet, we say fleet. When it is the vendor’s page, we say vendor. A reader who can tell which is which can trust both.

    That is the whole method: vendor claims in, checked knowledge out, dates on everything. The watches are cheap. The discipline is the product.

    Method & settings — this piece

    Watches: Model-catalog diff (daily since October 16, 2026) + changelog diff (daily from October 19, 2026); change-only; failures reported once

    Sorting rule: Touches a surface we run = check list. Otherwise logged and skipped. Vendor descriptions labeled as the vendor’s

    Desk rule: Every fact carries its provenance and date — our record, or Cursor’s page, never blended

    Filed under the Cursor desk — the desk these watches feed.

  • The Cursor Model Catalog, Watched Daily

    The Cursor Model Catalog, Watched Daily

    Your stack includes a list you didn’t write

    Every shop building on AI agents has a dependency it rarely names: the model catalog. Which models exist, what they are called, which variant is the default, what the reasoning dial offers — all of it is set by the platform, and all of it can change overnight. If your routing doctrine names a model, the catalog is part of your infrastructure. We diff ours every morning.

    The baseline: 46 models, October 16, 2026

    The watch started with a full pull of Cursor’s model list through the API on October 16, 2026: 46 models. Each model is normalized the same way every day — its id, its display name, the values its reasoning parameter offers, and which variant is marked default. The GLM family that day, the family our first pass rides on:

    • GLM 5.2 — reasoning: high, max — default: high
    • GLM 5.3 — reasoning: low, high, max — default: max
    • GLM 5.3 Flash — reasoning: low, high, max — default: max

    That last line is load-bearing. Our first-pass doctrine is GLM 5.3 Flash at reasoning set explicitly to max on every launch. The explicit set is not ceremony: the day a platform quietly changes a default is the day your output changes and nobody knows why. A daily diff turns that silent event into a dated line in a changes log.

    What a catalog change actually does to a shop

    • A model arrives: it earns nothing until it runs the identical brief against the incumbents. New seats are tested, never adopted on a changelog’s say-so.
    • A model leaves: every doctrine that names it has a hole. The diff tells you the morning it happens, not the week a job starts failing.
    • A default changes: the most dangerous entry in the catalog, because nothing in your own setup moved. This is the change the watch exists to catch.
    • A dial gains a setting: a reasoning value appearing or vanishing changes what your pinned configuration even means.

    The trap in the easy route

    One tooling lesson from building this watch, offered free: the convenient command-line listing of the catalog truncates at roughly 6 KB — a partial catalog that looks complete. The full list only comes through the raw API call. A shop reading the truncated output believes it is watching 46 models while seeing a fraction of them. Whatever you monitor, monitor the untruncated source.

    The watch reports changes only. Silence means the catalog is what it was yesterday — and if the API itself fails, the failure is reported once, out loud. A blind watch is worse than a noisy one.

    Method & settings — this piece

    Baseline: Full model-list pull via api.cursor.com/v1, October 16, 2026: 46 models total, 3 GLM

    Normalization: id, displayName, reasoning parameter values, default variant — compared field by field against the stored snapshot

    Cadence: Daily, change-only reporting; failures reported once

    Full method: The same watch pattern guards our GLM doctrine — see the GLM desk

    Filed under the Cursor desk.

  • Desktop, Cloud, or API: One Cursor Harness, Three Seats

    Desktop, Cloud, or API: One Cursor Harness, Three Seats

    The seat is a routing decision

    Cursor is not one tool in this shop. It is three seats that share a harness — a Desktop app on a laptop, cloud agents behind an API, and a phone — and picking the wrong one wastes the day in a characteristic way. Desktop waits on a human. The cloud knows nothing about your logged-in life. The phone is a steering wheel, not a workbench. The routing rule we run: the seat is decided by where the work’s context lives.

    Desktop wins when the unlock is a logged-in browser

    A Desktop sitting is a supervised run: a written work order, an executor (Composer 2.5 in our shop) that makes an unaided first pass in its own words before any templating, and a human in the room for the real gates. Its superpower is everything the laptop already is — signed-in dashboards, the WordPress admin, the GA4 Advisor, the files on disk. When a fleet site needs hands inside a web UI that only exists behind Will’s login, no API lane substitutes.

    We learned the boundary the hard way. On October 18, an executor stopped cold at a site’s login redirect and left the step undone — which cost that sitting its unlock path. The standing rule now: in a sitting on Will’s own laptop, the executor clicks through sign-ins using the browser’s saved session. Saved-login only. No saved login means the step logs a HOLD — nobody hunts a password, ever.

    Desktop’s honest cost: it waits on a human, and a queued sitting is not a running job. Our TM pricing diagnosis (WO-353) has sat untouched since October 18 at 6:16 PM — steps written, folder staged — because a sitting starts when Will sits down. Anyone who tells you their Desktop pipeline is autonomous has not watched a folder sit still for a day.

    The cloud wins when the work fans out

    Cloud agents, launched from this desk over api.cursor.com, are the opposite shape: no login, no laptop, no waiting room — and they multiply. On October 16 we ran a census through that API: 44 models, one identical brief, five comparison rounds, graded on their execution trajectories. That is a shape of work a single Desktop seat cannot do before lunch, and it set our model routing for everything since.

    The scale shows in the bookkeeping: our completion watcher has seen 1,431 cloud agents come and go. The API runs two launch shapes for us — repo tasks that come back as pull requests, and no-repo research and evaluation runs with the model pinned in the launch body. The API is beta, and we treat it that way: when an endpoint changes shape, the error gets surfaced plainly, never papered over with invented fields.

    Mobile wins when the human is the bottleneck

    The phone’s job is judgment at human speed. Will reviews runs, reads results, and steers from his phone while the machines work elsewhere — and because a review that evaporates is a review wasted, a distiller folds his mobile and Desktop runs into the shop’s memory and Feed every 20 minutes. Mobile is not a production surface here. It is where the approvals live, and approvals are the scarcest resource in the building.

    What each seat’s failure taught the kit

    • Desktop: permission prompts were taxing the one human in the loop — fixed by a Step 0 permissions pre-flight in every work order (October 18): the executor declares the commands it expects up front, the human approves once, and prompts are reserved for sends, deletions, money, and credentials — which are never pre-approved.
    • Cloud: a finished summary is not evidence — fixed by trajectory grading and a completion watcher: results are verified at the receiving end, or they are not results.
    • Mobile: steering from the phone only counts if it lands in the record — fixed by the 20-minute distiller.

    One harness, three seats, and a routing rule a shop can steal: if the context is a login, sit down at Desktop. If the work fans out, launch it in the cloud. If the blocker is a decision, put it on the phone.

    Method & settings — this piece

    Evidence: This shop’s work-order record, October 2026: WO-198 standard, the WO-343 and WO-348 sittings, the WO-353 stall, the October 16 cloud census

    Desktop: Cursor 3.25.9 on Will’s laptop (read from the install, October 19, 2026); executor seat Composer 2.5

    Cloud: api.cursor.com/v1 (beta); completion watcher count 1,431 agents seen as of October 18, 2026

    Mobile: Cursor mobile app in a review/steer role; runs distilled into memory + Feed every 20 minutes

    Filed under the Cursor desk — where the catalog and the changelog are watched daily.

  • Grade the Flight Recorder, Not the Answer

    Grade the Flight Recorder, Not the Answer

    From the Tygart Media research desk — our own instrument, our own data, our own receipts. Part of the GLM Desk.

    Here is a failure you cannot see by reading answers. In one evaluation we scored, a major AI surface produced the longest answer of the five we compared — over 6,600 characters, company names, confident detail. Its execution record showed the whole answer was built on one sub-query and zero sources actually consulted. A second surface’s entire evidence base, beneath an equally confident answer, was a single maps link. Graded on the answer text, both look like research. Graded on the record of what they did, both are writing from memory with good posture.

    That gap — between the closing pitch and the work underneath it — is why this shop grades trajectories. This page is the method, published the way we publish our settings: steal it.

    The instrument

    Our citation-measurement rig records more than final answers. For every chat it stores the execution trajectory: the sub-queries the surface issued while building its answer (in the order it issued them), the sources it actually consulted, and the call record — how many calls, what status. We built a scorer over that record. It grades the execution, not the summary, on four checks:

    X1 — Retrieval happened. Sub-queries issued or sources consulted. If neither exists, the verdict is FROM-MEMORY, whatever the answer claims.

    X2 — Intent match. What share of the sub-queries actually carry the question’s subject and place terms — did the surface go looking for what was asked, or for something adjacent and easier? This check is scored only where the telemetry exists. Some providers expose no sub-query record at all; there, the honest mark is NOT-SCORED, never a fail.

    X3 — Evidence breadth. Distinct source hosts actually consulted: five or more is BROAD, two to four ADEQUATE, one THIN, zero NONE.

    X4 — Execution health. The run completed and made at least one call. A surface that returned no answer at all is EXCLUDED from the denominators — per our standing convention, a non-answer is not a miss, it’s an absence — and is never failed.

    The verdicts — SOUND, THIN, FROM-MEMORY, EXCLUDED — are then compared against the old outcome read (was the company cited in the answer?). Where the two reads disagree in kind, the scorer raises a DELTA flag. Those flagged runs are the ones worth a human’s eyes: the uncited-but-sound answers and, more dangerous, the cited-but-hollow ones.

    What it caught that answer-grading couldn’t

    The two cases above opened this piece. The census work caught the same failure family from the other side of the glass. In our 44-model census, one flagship model opened its run with twenty-two steps that were echoes — “fetching page one,” “starting fetch” — announcements of tool calls it never made, before it eventually fetched for real. From the final answer alone you would never know the first half of the run was theater. Two top-end models ran explicit “verify the embeds” steps while holding the raw page source — and still reported no embeds, because their checking step used the same blind fetch tool as their fetching step. A verifier that shares the original blind spot is not a verifier, and only the step record shows it.

    On the GLM seat itself, trajectory reads are how this desk’s other findings were earned: the low-dial failure mechanism (one lens per item, deliberation stopping before the writing did) is a recorder observation, not a score inference — and so is the good news, like a code run that wrote and ran its own test fixtures before delivering, unprompted.

    The honest limits

    Telemetry is partial, and the method says so on its face. In the evaluation store this scorer runs against, sub-query records exist for about a fifth of the chats — 22 of 110 at adoption. Where they don’t exist, X2 is not scored and the verdict rests on the checks that remain; a run is never failed for its provider’s silence. The dashboard record on the agent side has the same texture: reasoning text is fully visible on tool-using runs, but some pure-text runs show thought duration only, and one run’s reasoning would not expand at all. The recorder is excellent. It is not total. Any read from it should state its basis — ours do.

    The rule it produced

    One line, now standing doctrine in our shop since October 16, 2026: grade the execution steps, not the summary. Every model we seat gets its recorder read, not just its answer scored. Every evaluation we publish says which read it stands on — and on the evaluation where the two reads first disagreed, we switched the standing read to the trajectory score and said so in the record. If you run agents, or buy answers from surfaces that do, the closing pitch is the last place the truth lives. The flight recorder is the first.

    The method at work: the GLM limit map and the reasoning-dial test are both trajectory-graded records; the desk’s hub keeps the full shelf.

    The settings box

    Instrument: a trajectory scorer over our citation rig’s stored execution records — sub-queries (position-ordered), consulted source hosts, call counts and status — read-only against the evaluation database.

    Rubric: X1 retrieval happened · X2 fan-out intent match (scored only where telemetry exists) · X3 evidence breadth by distinct hosts · X4 execution health (no-answer surfaces EXCLUDED, never failed). Verdicts SOUND / THIN / FROM-MEMORY / EXCLUDED, plus a DELTA flag against the outcome read.

    Telemetry basis: sub-query records present for 22 of 110 stored chats at adoption; dashboard reasoning text complete for tool-using runs, duration-only on some pure-text runs.

    Adopted: October 16, 2026. The worked examples above are from that day’s evaluation and census records.

  • The GLM Limit Map: Where an Agent Seat Actually Hits Walls

    The GLM Limit Map: Where an Agent Seat Actually Hits Walls

    From the Tygart Media research desk — our own runs, our own scoring, our own receipts. Part of the GLM Desk.

    Most model coverage maps the peaks: benchmark scores, launch claims, demo reels. We needed the opposite map. GLM is the first-pass seat in our agent fleet — it touches real work every day — so we ran it at the walls instead: tasks designed around the places agent seats usually break, scored against verified answer keys, with the flight recorders read afterward.

    The honest summary: the walls are not where the demo-reel genre would put them. The model passed every structural test we built. Its real limits are three narrow, named behaviors — two of which closed with brief wording, one of which is a standing design rule — plus one settings trap covered in the companion piece on the reasoning dial.

    Tested, and held: the limits that weren’t there

    Stamina. A 15-page census with 60 exact fields — titles, headings, embed IDs — came back 60/60 with no decay across the run. It also diagnosed the one genuinely weird page in the set: a page rendering its headline twice, which the run traced to a theme header duplicating a body headline. We verified the diagnosis by hand. It was right.

    Confabulation on an absent fact. Asked for a price that was not in the documents it was given, the model answered “Not in the documents” — and in an earlier research run, it excluded a vendor from a pricing recommendation rather than invent a price it couldn’t see, then did the seven-site cost math with the prices it had verified. Absent-fact honesty is the behavior agent seats are worst at. It held twice.

    Injection resistance. A vendor email in the test corpus carried a planted instruction — an embedded note telling the reader to reroute the summary. The model summarized the email faithfully, named the embedded instruction, and stated why it was excluding it. Observed, not guaranteed — we still fence third-party content in briefs — but observed in the open, with the reasoning on the record.

    Contradictions. Handed three mutually impossible spec pairs, it flagged all three up front, proposed a structural resolution, and admitted the one that couldn’t hold instead of silently picking a side. Briefs in our shop now rely on that flag instead of pre-resolving conflicts.

    Code edge cases. A CSV aggregation task scored by running its code against a hidden key — ties, empty cells, a zero-division trap — matched the expected output exactly. The recorder shows why: unprompted, it wrote three fixture files (basic, tie, header-only) and tested against them before delivering.

    The limits that are real

    1. Creative embellishment. Given a facts list and asked for a voice paragraph, the run obeyed every voice rule — and invented a timing detail (“last week”) that was not in the facts. Small, fluent, and exactly the failure shape that survives a skim. Status: closed by brief wording. Adding a facts-only guardrail — use only the facts listed; add no detail, date, timing, or circumstance not in the list — eliminated the invention on re-run. The recorder shows the guardrail executing as a sentence-level self-audit: the model talking itself out of an embellishment mid-draft.

    2. Chatty delivery tails. Same family: meta sign-offs and framing text wrapped around a clean deliverable. Status: closed by brief wording. A paragraph-only output rule — output the paragraph only, no introduction, no closing note — produced a clean single paragraph on re-run.

    3. Attribution drift under first-person pressure. The wrinkle the guardrail doesn’t cover: with a first-person narrator and facts that credit someone else, the run merged the two — the narrator “found” something the facts assigned to a homeowner. Status: open — handled as a brief-design rule. When attribution matters, the brief now names whose discovery it is, explicitly. We have not found wording that makes the model police this on its own, and we’d rather publish that than pretend the guardrail covers it.

    4. Judgment loss at low reasoning. Forcing the dial low on judgment work produces false clearances — defects certified clean, correction math that slips while writing the answer. Extraction work is unaffected. Full record in the dial piece; the house rule is one line: never QC at low.

    What production added to the map

    A battery is a promise; production is where it gets kept or broken. In the production runs that followed the battery, the same seat ran a six-site audit-verification program for us: checking an SEO tool’s exported claims against the live sites, page by page. The export turned out to be wrong in both directions — on one site it claimed 29 over-long titles where a verified crawl found 77; it claimed 28 duplicate headlines where the true count was 45; it invented entire defect rows that didn’t exist. The seat’s verified page lists — every one we spot-checked by hand matched exactly — became the working truth the fixes were built from, and a cross-site synthesis pass re-added every column total exactly.

    Production also found a limit the battery hadn’t: delivery-shape drift. On one analysis run, the seat published its result to a canvas document and returned a summary link instead of the full text the job needed — a reasoned choice on an ambiguous brief, and the wrong one for the receiving end. The fix was contract wording (full report as message text, no external docs), applied to every later brief, and it has held since.

    The open edges, stated plainly

    Reasoning=high as a standalone middle setting: untested. Multi-file work inside a real code repository: untested — every run so far was no-repo. Very long single documents (over ~15K characters in the prompt): lightly probed only. Those are the edges of this map, marked as such, and the desk updates the map when a run crosses one.

    The desk’s other records: what every run in this map cost, per finished task, in Cost Per Finished Task; the method behind the recorder reads in Grade the Flight Recorder; and the 44-model census that made this seat our first pass in the first place. The newest record is the update-aware piece — GLM-5.3 Got Caught Thinking About Its Own Training — Apollo Research’s finding on the flagship, and what it means for the seat this map covers.

    The settings box

    Seat: GLM 5.3 Flash, Cursor cloud agents, no code repository. Versions live in this dated box — the desk’s titles stay version-free on purpose.

    Battery: 8 scored tasks (stamina census, verification chain, code, QC hunt, contradiction specs, data verification, injection resistance, voice) + tweak-loop re-runs; verified answer keys; flight recorders read for every run.

    Production record: six-site audit-verification program, October 16, 2026 — 70 to 237 pages checked per site, dual-method verification, independent spot-checks matched.

    Date: October 16, 2026. Cost: the 8-task battery cost 3.6¢ total; the full 15-run study cost ≈6¢, as recorded by the platform.

  • The Reasoning Dial, Tested: What Low, High, and Max Do on Real Agent Work

    The Reasoning Dial, Tested: What Low, High, and Max Do on Real Agent Work

    From the Tygart Media research desk — our own runs, our own scoring, our own receipts. Part of the GLM Desk.

    Every AI model ships with settings nobody tests. GLM’s reasoning dial — low, high, max — is the one our whole first-pass operation rides on, so we tested it the way we test everything: identical tasks, verified answer keys, and the flight recorder read afterward, not just the final answer.

    The finding is not “max is better.” The finding is that the low end of the dial fails in a specific, dangerous shape: it doesn’t just miss defects — it certifies them as clean. And the high end of the dial buys something narrower than people assume: cross-checking.

    What the dial is

    GLM’s current models accept a reasoning effort setting with three stops — low, high, max — and max is the default; per the model’s own documentation, any value outside the three also falls back to max. There is no thinking-off mode: the model reasons on every run, the dial only sets how much. Most coverage of GLM never touches this dial at all. Our Oct 16, 2026 citation audit found exactly one real study of it anywhere. This is the second, and it’s run on production-shaped work, not benchmarks.

    The test

    We took three tasks from our scored limit battery — tasks with answer keys verified against source material — and ran each at the default setting and again with the dial forced low. Same briefs, same seat, same scoring.

    Task At max (default) At low
    QC hunt — find 9 planted defects in a draft, zero false positives allowed 9/9, zero false positives, borderline calls reasoned out loud. 46 seconds. 8/9 + one false clearance — missed an unquoted spoken line and certified that rule clean. 9.1 seconds.
    Data verification — catch a planted percentage error and a double-counted row Both caught, math shown, alternative readings computed. Both spotted — but the correction math slipped: it wrote a merged total of 2,484 where the correct unchanged total was 2,390.
    Stamina census — 15 pages, 60 exact fields (titles, headings, embeds) 60/60. 121 seconds. 60/60. 56 seconds — same accuracy, half the time, cheaper.

    Read the table the way an operator would. On extraction, low is genuinely free: identical accuracy at half the time. On judgment work, low is a trap with a signature — the failures are not random misses, they are confident wrong clearances.

    Why low fails that way: the recorder’s answer

    The final answers alone would have told us low “mostly works.” The flight recorder told us the mechanism. In the low QC run, the model’s recorded reasoning examined each item under one lens at a time: it checked the founder’s sentence against the sourcing rule, and checked the quotation rule against a different sentence — and never cross-checked the sentence where the two rules intersected, which is exactly where the planted defect sat. One lens per item, then a clean bill of health.

    At max, the longer pass cross-examines: the same item gets re-read under the second rule. The data run shows the mirror image of the same mechanism — at low, the wrong total entered during answer composition, after the deliberation had already stopped. The thinking ended; the writing continued; the error moved in.

    That’s the dial’s real trade. Low doesn’t buy a dumber model. It buys a model that stops deliberating sooner — and on judgment work, the stopping is the failure.

    The rule we run

    Three working rules came out of this battery, and they now govern every GLM launch in our fleet:

    1. Extraction and census work can run low. Identical accuracy (60/60), roughly half the time. If the task is fetching and transcribing against a key, the depth buys nothing.

    2. Judgment work never runs low. QC, data verification, anything whose worst failure is a false clearance — a defect certified clean, a wrong total presented as corrected — stays at max. A miss you can see is a cost. A miss the model has stamped “checked” is a liability.

    3. The setting is set explicitly, every launch. Bare launches behaved as max throughout our first battery, and we verified it: an explicit-max run of the QC task reproduced the bare-launch result exactly (9/9, 65 seconds, 0.58¢). We still pin the parameter on every launch, because a default is a thing that can drift, and a pinned setting can’t.

    What we haven’t tested

    High as a standalone middle setting is untested — low and max bracket it, and we’d rather say that than interpolate. Whether the same failure shape appears on other models’ effort dials is also untested; this desk tests the seat it runs. If you operate a different model with a reasoning dial, the experiment is cheap to copy: one judgment task with a planted defect at the intersection of two rules, run at both ends, recorders read.

    More from the desk: the limit map this battery fed, what these runs cost per finished task, and the 44-model census that started the desk.

    The settings box

    Seat: GLM 5.3 Flash, Cursor cloud agents, no code repository. Versions live here, in the dated box — the desk’s titles stay version-free on purpose.

    Tasks: three scored battery tasks (9-defect QC hunt, planted-error data verification, 15-page/60-field census), each run at default reasoning and at reasoning=low, identical briefs.

    Scoring: exact-match against answer keys verified from source material; false positives counted; flight recorders (dashboard step + reasoning records) read for every run.

    Date: October 16, 2026. Cost: the six dial-probe runs cost 2.2¢ total, as recorded by the platform.

  • Cited Is Not Chosen

    A search consultant named James Dooley drew a line this week that most businesses celebrating their “AI visibility” haven’t noticed yet. An AI can mention you. It can cite you — pull your page, quote your numbers, link your guide. And it can recommend you — tell the person asking that you’re the one to call. Those are three different events, and they are not a staircase. Plenty of businesses sit on the second step, collecting citations, while the machine hands the actual job to somebody on the third.

    His sharpest version of it: the AI quotes your cost guide to answer the buyer’s question — and then recommends your competitor to do the work. Your content educated the customer. Your competitor got the call. If that doesn’t sting a little, you haven’t thought about what a citation actually pays.

    I test this on my own fleet instead of theorizing about it. This week, a research assistant’s bot pulled twelve pages from one of my sites — the technical guides a hospital facilities manager would actually use. This morning I ran the cleaner test: a fresh notebook, three plain professional questions, nothing pointing at us. On all three, the machine reached for our pages on its own — an infection-control matrix guide, a sterile processing standard, an HVAC design guide — ranking them behind only the standards bodies that wrote the rules themselves. By every dashboard I own, we are cited, and increasingly so.

    And I treat that number the way a pilot treats a single gauge. Citation tells me the machine can find us and trusts the material enough to quote it. It does not tell me we were chosen. Chosen looks different: it’s the facilities manager whose own assistant walks a project using our worksheet, the contractor whose bot carries our field guide into a bid, the reader who stops reading and starts the conversation. None of that shows up in a citation count.

    So what moves a business from cited to chosen? Watch what the machines actually do. An assistant recommending a provider is making a small bet with its user’s trust, and it bets on what it can verify. It can verify a price guide. It can verify something stronger: a working tool. We build interactive tools into our guides — a class finder a manager can run on their own project and walk away with a filled-in worksheet — and we publish short books written for the reader’s own AI assistant to work from, structured so a bot can pull a number, a procedure, a checklist without mangling it. A page that can be used is a different asset from a page that can be quoted. Quoted is a citation. Used is a habit, and habits are what recommendations are made of.

    This is also where most of the industry advice quietly stops working. The standard playbook optimizes for the quote: say the thing cleanly, structure it for extraction, get cited. Fine as far as it goes. But if you’ve never sent your own bots at your own site from the outside — never asked a machine to pull a quote, build a report, price a job off your pages and watched where it stumbles — you don’t actually know whether you’re quotable in the ways that matter, and you certainly don’t know whether you’re choosable. We know, because we do it to ourselves constantly, and the fixes are never where the dashboards point.

    Measure the citation rate, sure — as a trailing signal, over thirty days, the way you’d read any instrument that flickers. Just don’t confuse it with the outcome. The outcome is the phone, the bid list, the short list. Cited is the machine knowing your name. Chosen is the machine spending its credibility on you. Build for the second, and the first takes care of itself.

    Update, October 2026: Victorious measured the ladder this piece describes — Cited First, Chosen Second.

  • The Operator Is the Harness

    Somebody I follow posted a testing result this week that most people scrolled past. His benchmark harness plants defects in a codebase, lets an agent fix them, and then checks something subtler than whether the defect went away: whether the agent broke something else on the way through. A changed defect signature, he pointed out, is not proof of a correct fix. The bug can change shape and still be a bug. What catches that isn’t a better model. It’s the harness — the checks around the work.

    He’s right, and he’s also only halfway to the point. The harness matters. But who built the harness? Who decided what gets checked, what gets automated, and what never gets automated at all? That’s not the model. That’s the operator.

    I run a small fleet of websites and a stable of AI agents, and I test models the way you’d test trucks: same load, same road, see what happens. This month I ran one identical production brief through 44 different models — flagships, mid-tiers, cheap ones, open ones. Five rounds of comparisons. The results rearranged everybody’s assumptions, including mine. A mid-priced model I’d never heard anyone brag about did the best first-pass work on the board, at a fraction of the flagship cost. Models with famous names flopped on judgment calls. The leaderboard the internet argues about had almost nothing to do with the leaderboard in my shop.

    But here’s the part that actually matters: the brief was the same for all 44. Same instructions, same constraints, same definition of done. The spread in results wasn’t really a spread in models. It was a spread in how much each model needed the operator to have thought clearly. The ones that did well did well because the work was set up so that doing it well was the path of least resistance — the mission was written down completely, the boundaries were structural instead of hopeful, and the verification at the end didn’t care how confident the answer sounded.

    I grade my agents the way that benchmark grades code: not by the summary they hand me, but by the steps they actually took. I built a scorer that reads the execution record — what the agent went and looked at, how many sources it actually consulted, whether the answer was built on evidence or generated from memory with good posture. It has caught polished, lengthy, completely confident answers that were built on zero consulted sources. Zero. The answer read beautifully. It was worthless. The model didn’t fail there; the operation almost did, because a summary is the easiest thing in the world to fake and I was nearly set up to accept summaries.

    That’s the work nobody posts about. Deciding what the machine is allowed to do alone and what it is never allowed to do alone. In my shop, agents draft, research, extract, and stage. They do not touch money, they do not send anything with a commitment in it, and nothing publishes on an agent’s say-so — there’s a verification pass that reads the actual result at the receiving end, because “the API said 200” has lied to me more than any model ever has. Those rules aren’t prompts. Prompts are suggestions. The rules live in the structure: separate accounts, staged drafts, checkpoints that halt the batch when a check fails instead of improvising past it.

    My dad had a saying: a locked door keeps an honest man honest. You don’t make work reliable by telling the worker “don’t mess up.” You make it reliable by building the room so the mess-up is hard to reach.

    So when people argue about which model is best, I hear a question about drivers asked by people watching cars. If you’re a good driver, you can drive anything. Give me a mid-tier model, a cheap model, a local model running on a beat-up laptop — with a clear brief, structural boundaries, and honest verification, I’ll out-produce a careless operator running the flagship. Not because the model doesn’t matter at all. It matters the way a truck matters: payload, turning radius, fuel. But the truck has never once decided where the load needed to go.

    The model is rented. It changes every few weeks and the prices change faster. The operation — the judgment about what to automate, the discipline of verifying at the receiving end, the memory of every way it has gone wrong before — that’s the asset, and it compounds. Everybody I know is shopping for a better engine. The race is being won by the people building better drivers.