agent-ops-patterns / gallery
Who decides what inside open-source agents
Each map takes one open-source agent system and lists the recurring judgments it found there, each as a seat: the question the judgment answers and who holds it (code, a model, or a person), with how it has failed and what checks it where the mapping found them. Seats are sorted rule vs ruling and human vs machine. Seats a model holds identify the prompt pieces and assembly paths found in the mapped source. The OpenHands map also quotes prompt source text on 14 of its cards, each quote linked to the lines it came from. Each map has a report of the gaps the mapping exposed.
The method is "Cognitive Operations Maps".
The maps
OpenHands/software-agent-sdk, the agent loop of the OpenHands coding agentIts goal judge reads "complete": "false" as complete, because in Python bool("false") is True. We reproduced it with a failing test and reported it as #5492, which the project has labelled ready-for-dev. Two outside contributors have since opened fix PRs, #5493 and #5563 (one a draft); neither is merged yet (checked 2026-10-12). All the gaps →
aae9c437, 2026-10-06Make a map of your own system
The skill that builds one is skills/cog-ops-map/. In Claude Code:
claude plugin marketplace add u00dxk2/agent-ops-patterns
claude plugin install agent-ops-skills@agent-ops-patternsOr copy the folder; it is in the Agent Skills format.
How to read one
Each map is one self-contained HTML file with no external requests. The links above open it in your browser; the files are also in the repository's gallery/ folder if you would rather download one. A "Download .md" button on the page exports the seat cards and the gap report as Markdown, if you'd rather hand it to your own assistant.
What a map is, and is not
- Not exhaustive. Each map names the folders it covered and how its seats were found; a judgment outside those, or in a helper nobody opened, is not on it.
- A snapshot. It describes the code at the commit it names. The project keeps moving, so a cited line number can stop pointing at what it describes; each map says which parts were checked at which commit.
- A reading, mostly. Each gap says whether it was reproduced with a test or read from the source. Most are readings of code paths, not observed failures.
- Not a security audit, and not a verdict on the project. The maps describe behaviour. Where a project's own docstring states a choice, the map quotes or cites it.
- Shared with the project first. Findings that were reproduced are reported to the project in public before or alongside the map, and the map names those reports by issue number.
Corrections are welcome as an issue. A wrong sentence about someone else's code is the worst bug a map can have.