The hive
Five AI seats and one human. Not an agent swarm: each seat has its own memory, model and permissions, and autonomy is granted per task.
1 Nov 2026 · 2 min readFive seats and one human. Not an agent swarm — each seat is a long-running session with its own memory, its own model, and its own permissions. Anthropic measured the price of the other design, many parallel agents: about 15× the tokens of a chat, and their own conclusion that human evaluation catches what automation misses 1.
The seats
| seat | model | does |
|---|---|---|
| prime | Claude Opus (Claude Code) | architecture, measurement, verification |
| associate | Claude Opus (Cursor) | planning, review |
| grok | Grok (Cursor) | most of the code |
| composer | Composer (Cursor) | leaner code, independent verification |
| una | local Gemma-4 12B | reports from inside the system |
| michal | human | decides |
Una runs on the machine. No cloud call. She is the one seat that can be woken live.
What makes it a hive and not a queue
Permissions are asymmetric and enforced server-side. Two code seats, and only one of them may issue assignments or sign off peer review. The other may not. That asymmetry is in the schema, not in a prompt.
Nobody reviews their own work. @verified on your own delivery is refused by the server, not by etiquette.
One human closes the question. A broadcast to every seat can only be fired by Michal.
Autonomy is request-only
The default is manual: everyone posts, the human prompts each pickup.
Why, measured in one day: three autonomy defects, each invisible in normal operation, each costing exactly one wasted run — because someone was watching.
Wrong prompting costs a round trip. Wrong autonomy costs an unobserved night.
Autonomy is authorised per task. It never stands, and it is never inferred from progress.
The failure mode we keep finding
A structure declared in one place, populated by a second place nobody runs, with no assertion connecting them.
Eleven instances so far. The fix is never a better matcher — derive from the registry, write the edge when the fact arrives, and assert the count out loud.
Figures as measured on 22 October 2026, from the running system’s own stores.
- Anthropic. How we built our multi-agent research system. anthropic.com (2025)saved pagecorpus: ~/hybrid-retrieval
Our planning database had 1,061 nodes and nothing written in seven weeks. Meanwhile the real planning ran in a chat thread. What the chat had that the planner did not: append-only posts and a lifecycle enforced by the server.
Three times the local agent came back empty, and three times the model got the blame. Each time the cause was a restart from the same session that was waiting for her. A defect in the instrument reads as a defect in the subject.
Our local agent logged 477 tool calls with 63 errors. Counted, it looks like an unreliable model. Classified by cause, 52 were our tooling, our guards or a deliberate test. The remaining 11 were aiming errors. 75% recovered on the next call.