The coordination problem
Two agents grab the same task and send a prospect two emails. One agent finishes half a job and nobody picks up the rest. Coordination failures like these are the most common way multi-agent systems embarrass themselves.
Typed handoffs
A handoff is a small, structured note: what the task is, what's done, what's left, which sources matter, and who it's for. It's claimed by exactly one agent, which holds a lease on it until it finishes or gives it back. If the lease expires, the task returns to the queue. ai-memory uses this pattern for handing work between coding agents.
| Field | Example |
|---|---|
| Task | Book calls with the 12 accounts that replied yes |
| From → to | gtm-chief → chief-of-staff queue |
| Done so far | Replies logged, accounts updated in the brain |
| Sources | Reply thread IDs, account pages |
| Lease | Claimed by coo, expires in 2 hours |
Read-only verifiers
The agent that did the work never grades its own work.
Every important output gets checked by a different agent with read access only. It compares the output against the brief, the sources and the claim library, then approves or sends it back with a reason. Because it can't edit, it can't quietly fix a problem and hide it. COG Second Brain uses read-only verifiers the same way.
Where humans come in
Verifiers catch mistakes. Humans make judgment calls. The chief-of-staff routes anything the charter marks as high-stakes, like pricing, contracts and regulated advice, to the founder or the responsible professional, with the handoff, the sources and the verifier's notes attached.
The audit trail
Handoffs, leases, verifier decisions and approvals are all logged. When something goes wrong, you can see who did what, with which sources, and who approved it. That's the difference between "the AI did it" and an answer you can act on.
Building in a legacy industry?
We bring the agent crew and the distribution. You bring what you know about the industry.