Self-governing agent organizations.
Apiary is a substrate for building agent teams that govern themselves. It borrows from three places at once: how brains route work between specialized regions, how legislatures debate and vote, and how good founders oversee rather than micromanage.
The brain-region model
Most agent frameworks are flat — a swarm of identical workers — or hard-coded as rule-based bots. Apiary treats every capability as a specialized region with its own folder, its own README, its own failure mode.
When one region degrades, the rest keep running. When work arrives, the thalamus routes it to the right specialist. The system is resilient because it is decomposed, and efficient because no region does work it isn't built for.
The committee model
Every sector — billing, ops, UI, safety — runs as a committee of 3–5 reps. Reps draft proposals. The committee debates and votes. The elected CEO breaks ties, in writing. Only the genuinely novel, ethical, or irreversible decisions reach the founder.
Decisions are provisional. Any rep can move to reopen a settled decision — unanimous to reopen, majority to amend. The hive learns by compound revision.
The cortex map
Eight brain regions, each a real folder under brain/, each with a README declaring its identity. The directory tree is the literal nervous system.
Planning, decisions, working memory — the orchestrator.
Long-term memory — .md persistence across sessions.
New memory formation — deciding what's worth saving.
Pattern recognition, screenshots, OCR.
Prediction — next-step inference from past sequences.
Safety — threat detection and agent containment.
Procedural memory — learned habits and fast paths.
Input router — dispatches each request to the right region.
The six pillars
The whole architecture decompresses from six words.
Cortex
Brain-region modular architecture.
Each capability is a specialized region (a folder, a substrate, a role). One region down does not mean the system is down.
Committee
House-of-Reps governance per sector.
3–5 reps per sector propose, debate, vote. CEO breaks ties. Founder only sees the genuinely novel.
Cater
Symbiotic developer-agent loop.
Folder = lobe. File = memory. README = identity. The structure you give the agent is the brain it inhabits.
Reconsider
Self-correction via revote.
Decisions are provisional. Any rep can move to reopen. Unanimous to reopen, majority to amend. Compound improvement over time.
Oversee
Founder as constitutional overseer.
Read-only visibility into every committee decision. The option to intervene never disappears; the need does.
Situational
Felt-sense judgment over rule-following.
Agents detect when the work is done by outcome, not by retry counts. Calibrated by persistent memory.
Why this, why now
- ▸Bring your own LLM. No baked-in keys, no token surprises, no provider lock-in. Switch providers anytime.
- ▸Founder-overseen. Read-only oversight of every decision. Intervene when something is genuinely novel — otherwise, let the system run.
- ▸Self-improving. Decisions are provisional. Old processes get retired via the same voting flow that adopts new ones. Compound improvement built in.