Agent Ops Patterns

Overview

Agent Ops Patterns is a public gallery of cognitive operations maps. Each map takes one open-source agent system and lists its recurring judgments as seats: the question each one answers, who holds it (code, a model, or a person), how it has failed, and what checks it. Where a model holds a seat, the map identifies the prompt pieces behind it, and every map ends with a report of the gaps the mapping exposed.

The first map covers the OpenHands agent SDK: 171 seats and 70 gap entries, all read at one commit from 2026-10-06. Gaps that were reproduced with tests went to the project as public issues first. Each map is one self-contained HTML file you can read in the browser or export as Markdown.

The same repository holds the skill that maps your own system, a five-question audit you paste to your agent, and small tested libraries. A map is a reading of code at one commit, not a security audit.

Why I built it

I run my software portfolio through a lot of AI agents, and the question I kept failing to answer was simple: which decisions is a model making, and what catches it when it's wrong? Org charts and process docs don't say where the judgment sits. So I started writing it down, seat by seat - first for my own systems, then for open-source agents anyone can check.

How it works

1. Open a map - start from the gallery; each map opens in your browser as one page 2. Read the seats - each card names the question, who holds it, how it has failed and what checks it 3. Check the gaps - each entry says whether it was reproduced with a test or read from the source 4. Map your own - install the skill in Claude Code and ask your agent to map a system you work on

Feature highlights

• Seat cards - one card per recurring judgment, sorted rule vs ruling and human vs machine.
• The prompts behind the seats - model-held seats point to their prompt pieces; 14 OpenHands cards quote the source text, linked to its lines.
• A gap report - what the mapping exposed, each entry marked reproduced or read from source.
• Findings shared first - reproduced gaps are reported to the project in public, and the map names the issues.
• Markdown export - hand a whole map to your own assistant.
• Map your own - the cog-ops-map skill installs as a Claude Code plugin.
• Five-question self-audit - paste one prompt to your agent and it scores your setup out of five, read-only.

Pricing

Free and open source under the MIT license.

Notes

• For whom: People building or running AI agent systems, and teammates who need to see where the judgment sits without reading the code.
• How it's built: Static pages on GitHub Pages; each map is a single HTML file that makes no external requests.