This week's collection adds real accounts to a directory that began with illustrative examples. The accounts span different publication dates. They are newly collected here on October 14, 2026; they are not all events from the past seven days. We selected them for concrete work, visible artifacts and evidence about the human role.
Read each entry as an attributed account. A repository, session trace or research paper can help you inspect a claim, but inclusion does not mean we reproduced the author's deployment. The collection also contains outcomes that remained incomplete or failed. Those accounts are useful when you need to decide what an agent may do next.
A bounded patch has a clear result
The Codex blog repair began with a visible formatting defect and a request for a test. It shows why a small, familiar repository problem can make a useful first delegation.
Our editorial lesson is to choose a check that points to the behavior the user sees. If a social preview is the problem, inspect the preview. If a page throws an error, visit the page. Passing commands support the judgment, but the relevant user result still needs attention.
The collection includes several much larger coding tasks. That does not make the smaller task obsolete. A small repair lets you assess environment setup, task interpretation and review effort with less ambiguity. It also gives you a recoverable artifact to compare against your own implementation.
Compatibility suites make large tasks judgeable
In the JustHTML port, Willison supplied an existing parser and an independent compatibility suite. Codex worked through a reviewed specification and repeated checks. The account includes skipped cases, which matter when interpreting the result.
A port is a useful example of a task with a substantial amount of work and a relatively concrete target. You can compare behavior against a suite maintained outside the new implementation. That is stronger evidence than a test the agent writes solely to describe its own choices.
Before borrowing this pattern, inspect licensing, unsupported behavior and the suitability of the suite for your intended use. A parser that passes compatibility tests may still need security review, performance evaluation and maintenance ownership. The directory's reported outcome field deliberately avoids turning test success into a promise of production readiness.
Browser access supplies another way to check work
The GIF optimizer account connects a coding goal with browser verification. Claude Code used Rodney to exercise its interface and repair observed UI problems. The source also records follow-up requirements about the build and attribution.
Use that pattern when your artifact is interactive. Give the agent a representative file and an expected sequence of actions. Ask it to operate the controls, inspect results and preserve evidence of what happened. A screenshot of the initial page does not establish that upload, settings and download work together.
The presentation recording entry covers a different browser task: operating a local page to produce a clip. Its output is directly inspectable. That kind of bounded browser operation can be easier to evaluate than a task spread across several authenticated business services.
Operational work can expose design errors
Dodds' hybrid-search account includes a production failure after an agent-built implementation. Foreground index maintenance blocked requests. Later redesign moved search into a unified Worker.
Treat that account as a reminder to evaluate architecture under the conditions where users experience it. Ask who owns derived data, when synchronization happens and whether a cold request can block unrelated work. Local relevance checks can pass while the runtime design remains unsuitable.
For your own agent task, include an operational question in the completion contract. For example, require a report on what happens when an index is missing or a downstream service fails. Keep the question specific to the system. A long generic checklist can hide the one failure that actually matters.
A useful research run can remain unpublishable
Willison's orchestrion research experiment returned structured location records through repeated tool use. He explicitly considered the list incomplete and in need of fact-checking. The existence of a JSON artifact did not resolve the quality of its contents.
The existence of a JSON artifact did not resolve the quality of its contents.
That distinction belongs in every research use case. Separate retrieval success, useful leads and verified conclusions. Preserve source URLs at the claim level when possible. Record missing evidence and contradictions in the same artifact as the answer, so a later reader can see the limits without opening another system.
A prompt asking for every instance of something creates a difficult completeness claim. Narrow the scope by geography, date, accessible corpus or named institutions. If you cannot define a test for completeness, label the result as a collection of discovered records and document the search process.
Scientific agents still have human execution boundaries
FutureHouse's Robin report describes a loop linking literature research, experimental proposals and analysis. Human scientists carried out the physical experiments. The reported therapeutic candidate remains a preclinical result.
The lesson for a research team is to name who controls each transition. A computational analysis can suggest the next experiment without being authorized to carry it out. An experimental result can support a hypothesis without establishing a clinical recommendation. Keep those distinctions visible in the use case record.
If you adapt the pattern, start with a supplied dataset and a narrow analysis question. Ask the agent to produce reproducible code, assumptions and proposed follow-up work. Have a qualified researcher review the methods and interpretation before deciding on new experiments.
Failure accounts deserve space in the directory
Project Vend's first report describes a real shopkeeping experiment that lost money. The agent had tools and ongoing authority, yet made poor commercial decisions. The second phase added capabilities while preserving human checks on purchases.
A useful catalog must make room for that kind of evidence. Counting entries alone rewards optimistic summaries and obscures unsuccessful runs. We want a reader to find the constraints alongside the task and understand what the author actually observed.
Browse the agent guides to compare accounts using the same system. Read the primary source before adapting a prompt, and define your own completion check. Subscribe through RSS for newly published entries and field guides; the feed preserves example labels where video previews remain illustrative.