Multi-Agent Orchestration

These examples and illustrative results are independently authored teaching materials, not measured model results.

Use case

Review API correctness and accessibility on one candidate, with relatively independent scopes. Teaching agents are read-only and a coordinator collects/verifies/merges. Agreement is not truth by vote.

Mechanism

Check coordination value, assign bounded roles/common revision/necessary inputs/report contract and collect actual results. Merge by issue identity with evidence while preserving distinct triggers/disagreement. Any editing needs explicit ownership/integration rather than concurrent uncoordinated changes.

Bad example

Spawn many all-scope reviewers, accept the majority and declare completion before collection, or have everyone edit the same files.

Good example

Assign teaching A API and B accessibility read-only at one fixed revision. Require locations, triggers, evidence and uncertainty. The coordinator collects/verifies and merges genuine duplicates, retaining disagreements without vote-based truth. Unfinished outputs mean incomplete review. Separate implementation ownership/integration checks from review.

Why the change matters

Roles plus revision make results comparable. Collection/verification converts concurrent reads into delivery rather than treating spawned tasks or consensus as completion.

Observable expectation

An illustrative report includes both actual states/revision/evidence and disagreement; unfinished work stays visible.

Inspect evidence-supported acceptance, preserved duplicate provenance and no edit collisions. No real team is started here.

Limits

Use one agent when work is not independent or coordination outweighs it. Source context-isolation rules are workflow choices under current host/authorization. More agents promise no quality/coverage guarantee.