When Codex, Cursor, Claude, and Spot all touch the work, the risk is not just slowness. It is summaries becoming facts and stale material becoming current truth.
02 / Case File
TITI Central Agent
Giving multiple agents a boundary-aware operating order
A smarter agent system is not one that remembers more. It knows what not to remember and when to return to original evidence.
Live proofAuditable boundaries around memory, review, handoff, and agent collaboration.
Separate generated briefs, digests, review aids, and original evidence. Turn handoff, training samples, durable memory, and human confirmation into collaboration protocol.
Give Perri multi-agent speed while preserving evidence discipline and business judgment.
When Codex, Cursor, Claude, and Spot all touch the work, the risk is not just slowness. It is summaries becoming facts and stale material becoming current truth.
Separate generated briefs, digests, review aids, and original evidence. Turn handoff, training samples, durable memory, and human confirmation into collaboration protocol.
Give Perri multi-agent speed while preserving evidence discipline and business judgment.
02 / Evidence checklist
Evidence checklist
Durable memory admission rules
Handoff prompt templates
Review-aids-only principle
Agent collaboration map
Workflow
Workflow
Identify source type
Original evidence, summary, draft, and review aid are treated differently.
Route roles
Spot finds, Codex builds, Cursor executes crisp tasks, Claude explores.
Keep human review
Durable memory, client material, production release, and business conclusions return to Perri.
Guardrails
Guardrails
- A summary does not become durable memory by itself.
- Token saving is a useful side effect, not the primary goal.
- Dynamic operating data must pass freshness gates.
What it proves
What it proves
- Multi-agent work behaves more like a team than a chat pile.
- Handoffs lose less context.
- Mistakes enter review instead of becoming the next hallucination.
Next motion