Patterns

Evergreen pages on the agent-orchestration patterns the lab runs. Each one opens with its drawing.

1. Adversarial reviewstandardReview the work with models of different families, one told to break it, and back the verdict with deterministic checks.

2. Deterministic gatesstandardShell checks decide pass or fail, never an LLM judge, and a failure carries its reason into the next prompt.

3. Cross-platform checkin useThe reviewer runs on a different vendor from the implementer, so one family's blind spots and one quota ceiling cannot stall the graph.

4. Deterministic commitin useA check node commits with the spec's own Conventional Commit subject, and a model never writes the commit.

5. Event intakein useA graph that fires on a watched file, turning outside events into specs while nobody is watching.

6. Parallel proposersin useSeveral cheap models draft the same spec in parallel, then one implementer builds on the best ideas.

7. Resilient membersin useCascade and round-robin ensembles give one node a fallback ladder or a rotation across providers, so a quota death is not a failed run.

8. Reviewer fixesin useThe reviewer who rejects the work fixes it, instead of writing change requests for another model to read.

9. Review panelexperimentAn experiment, each reviewer owns one axis, security, performance, maintainability, or project patterns, and every axis must pass.

10. Task routerexperimentAn experiment, a router judges the task by kind or complexity and sends it to a model suited to it.