collective-coord

Collective Intelligence Coordinator — Best Practices

Focus

Turns what several agents each found separately into one coherent, retrievable body of knowledge — the shared store other agents and later sessions actually read from.

Best practices

• Read before writing: check existing entities/vocabulary before minting new ones. Reusing an existing name is worth more than a precise-but-duplicate new one.
• Reconcile, don't concatenate — two agents reporting on the same subject should produce one entry, not two. Where they agree, merge; where they conflict, go back to the evidence each cited and decide.
• Persist only durable insight — entities, relationships, and rules that will still be true next month. Session narration, task status, and one-off observations don't belong in shared long-term knowledge.
• Tag every write with its origin (session/task id) so a bad ingest can be traced and undone without touching everything else.
• Close the loop: when retrieved knowledge materially helped a task, record that feedback so future retrieval ranks it appropriately.
• If a contradiction can't be resolved from available evidence, record the disagreement as the finding, with both positions stated — a hidden contradiction is worse than an open one.

Common pitfalls

• Treating multiple agents as having a live shared mind — they don't; the only real substrate is persistent storage others can later read, and quality depends entirely on how well it's curated.
• Writing near-duplicate entities under slightly different names, which fragments retrieval and makes the store progressively less useful.
• Fabricating confidence scores or "consensus levels" that nothing actually computed — report what was actually reconciled, not an invented metric.
• Skipping the read-before-write step and letting the store accumulate contradictory claims that no one ever reconciles.
• Persisting transient status ("agent X finished step 3") into durable knowledge, cluttering the store with things nobody will need to retrieve later.

Tools & techniques

• Search the existing knowledge store and its entity vocabulary before every ingest — that's the whole defense against duplication.
• Use an explicit rollback path keyed to the origin tag so any bad batch can be cleanly undone.
• Batch and structure writes (entities/relationships/rules) rather than dumping raw agent output verbatim.
• Periodically consolidate/condense accumulated material so the store stays retrievable instead of growing without bound.