Self-hosted · any model · set up by your coding CLI
Benzene runs AI agents as cells. Each cell walks a six-atom ring for every instruction, and a compiler, a test or a document check decides whether the work counts. Cells that keep failing divide. A colony grows into an organisation of six offices, founded in order by its own record, and a heartbeat keeps it running itself.
Pick the CLI, pass the purpose. Claude Code, OpenCode or Codex does the rest: installs, picks your model, writes the manifests, founds the organisation, runs the first check.
the ring one instruction, six atoms, in order · light where the work is now · CEO the seat, not a cluster
From one purpose to an organisation
One sentence becomes one cell. The cell heals against its guardrail, divides when it falls short, founds offices as the product matures, and evolves on analytical feedback until the purpose is delivered. Watch it happen.
One cell, six atoms
The ring is the whole company inside one cell. At the general scale each atom is an office of many cells; at the universal scale the same six govern galaxies. The titles change with the rung of the ladder. The ring does not.
Every cell states the problem, plans the steps, runs them alone or one session per skill in parallel waves, and writes an honest reflection after the verdict is in.
Skills are SKILL.md files any model can follow. Tools write to the workspace, ask an operator, search the web. Verified executions become new skills.
Anthropic, OpenRouter, a local server. Roles per model: primary, judge, router, evolver, scout, and a decider that routes each instruction to the right office and cell. Free tiers survive outages with fallbacks and retries.
Compilers, tests, import resolvers, document structure, claims held to the record, a balance sheet that must add up. A failed verifier caps the score and heals the cell.
Guardrail efficacy and world efficacy, each an EMA. The guardrail cannot vouch for work the world rejects, and the world cannot excuse work that fails the guardrail.
A task is offered to humans only after every verifier passed. The rest stays in the heal loop. Evaluation panels judge the record like a Product Hunt or a Shark Tank.
Work that has to pass
Other frameworks approve agents. Benzene verifies their work. Ten verifier types run the real toolchain after every attempt, and the framework goes further than hand-written checks:
# the CTO's genome: the product's own checks
guardrail:
verifiers:
- {type: command, name: verify:cli, command: [node, --check, index.js]}
- {type: sections, files: [skills/*/SKILL.md],
front_matter: [name, description]}
require_written_files: true
pass_threshold: 0.70
# what the scout added, proven on a spoiled copy
guardrail_source: scout:scout_e5d88655
guardrail_evidence:
- verify:skill-doc: passed clean in 0.1s; failed spoiled
The heartbeat
Once the offices have passed their guardrails, nobody has to type instructions. Every office does its duty on its beat, in founding order, and everything flows through the COO. Features wait in the CEO's inbox. Maintenance just happens.
| Office | On its beat | To the CEO |
|---|---|---|
| CTO | runs its verifiers on the workspace, fixes what fails, releases | maintenance done, or needs you |
| CXO | convenes its panel, studies the experience and the competition, suggests a feature | |
| CFO | reads the balance sheet and the growth curve, suggests a change | when the system burns |
| CMO | reads the CXO's analytics, builds the campaign, asks for product work | |
| CIO | cheap panel first, the angel panel when it says yes | the offers |
| COO | researches the beat, proposes, builds what is approved, measures before and after | approvals, shipped, improvement reports |
Where it sits
One self-hosted agent with memory and skills, reachable on chat. Grows with you. Nobody checks its work but you.
The Hermes shape.
Cells with guardrails that prove the work, an organisation founded by its own record, a heartbeat with a CEO inbox. Self-hosted, one process, any model.
Proof and governance in the same loop.
Hire any agent into a role, wake it on a schedule, approve at board level, cap its budget. Governance without a verifier.
The Paperclip shape.
Quick start
benzene plugin start --for claude "build me the best component library". Or --for opencode, --for codex. The skill lands where that CLI reads skills and the CLI opens with the request.
It installs what is missing, picks the model from the keys you have, reads the project, writes benzene.yaml and benzene-org.yaml with your purpose, runs the first check and reports which offices are founded. At most two questions.
Your original stays as written. A cluster block under it records how the cell healed, evolved and divided.
/benzene:heartbeat and /benzene:inbox, or bza --heartbeat --beats 0 --every 3600. Features wait for you; maintenance runs on its own.
$ benzene plugin start --for claude "build me the best component library"
installed the benzene skill at ~/.claude/skills/benzene · handed over to claude
OPENROUTER_API_KEY found · wrote benzene.yaml (verify:test, verify:types) and benzene-org.yaml
$ bza
=== round 1: work discovered by verify:types; 1 task(s) by skill efficacy
[CTO] cto took: fix the type error in src/header.tsx ... passed 0.93 verified
office coo founded: the CTO cluster is healed
$ bza -i "the CLI crashes on init, fix it"
[CTO] cto took: the CLI crashes on init, fix it
routed by decider (1.00): typesafe/jev-router
$ bza --heartbeat
=== beat 1
cto ran verifiers pass: verify:cli, verify:skills
coo ran the CXO and CFO both point at dark mode
COO proposes [feature] Dark mode (pending): approve with --approve prop_f27dde47
Run your own organisation of agents. Self-hosted, open source, verified at every step.