The new era is already born.
Claude Code, Codex, OpenCode, Antigravity and every model behind them now live on your machine. Intentocracy governs them as one organization. You state the intent. They receive mandates.
The crowd
Every machine now hosts a crowd.
Each one says it can do the job.
Nobody decides which one should.
Until now.
0 agents, models and sessions open at once
Every system of government answered one question: who may act, and on whose behalf?
We read four thousand years of answers and kept one rule from each.
-
Code of Hammurabi
Babylon carved its law in stone before any case was heard, so no judge could invent it afterwards.
We keptA constitution agents cannot rewrite.
-
The Roman Republic
Two consuls, elected for one year, each able to check the other. No office was permanent.
We keptEvery mandate is scoped and expires.
-
Athens, after Cleisthenes
The assembly of citizens held power. Officials were drawn by lot and answered to it.
We keptThe human is sovereign. Agents serve the stated intent.
-
Sui imperial examinations
Office went to whoever passed the examination, not to whoever claimed the right to it.
We keptWork is routed by measured merit, never by a model's claim.
-
Magna Carta
For the first time the king himself was bound by written limits he could not lift.
We keptLeast privilege, enforced outside the model.
-
The Venetian doge election
Ten rounds of ballots and lots, so no family could capture the choice.
We keptIndependent verification for critical work.
-
The Ottoman Divan
The imperial council copied every decree into the mühimme registers, so any order could be traced.
We keptEvery decision is recorded and explained in one line.
-
The separation of powers
The one who makes the law, the one who carries it out and the one who judges it are never the same.
We keptExecution, verification and acceptance are separate.
-
The Swiss federation
Power stays with the smallest level that can hold it. The centre gets only what it must.
We keptLocal-first. Your history lives on your machine.
-
Intentocracy
The first government whose officials are agents. Its citizens state intent; nobody picks a model by hand again.
RuleHuman intent governs. Agents receive mandates.
One sentence becomes a mandate.
- Intent
- Classify
- Route
- Mandate
- Execute
- Verify
- Record
An illustrative run. Field names, agents and checks are the ones the control plane uses today.
you@mac ~/shop-web $ intentocracy "make the orders page faster"
intent persisted int_7f3a2c
project resolved shop-web (git root, npm)
capabilities discovered 14 agents · 4 runtimes · 9 healthy
You never named an agent, a model or a test.
Classification
- bucket
- code_repo_change
- risk
- medium rules: touches request path
- size
- S complexity 2
- privileged
- none no git push, network, secrets
- decided by
- deterministic rules → verified history → Ollaya (only above 0.6 confidence)
Routing expected cost per accepted outcome
| executor | p accept | attempts | cost / accepted |
|---|---|---|---|
| agy | 0.91 | 1.10 | 35k tok |
| claude:default@haiku | 0.78 | 1.28 | 61k tok |
| codex:default@gpt-5.6-luna | 0.83 | 1.20 | 118k tok |
| claude:default@opus | 0.88 | 1.14 | 335k tok |
| opencode:build@big-pickle | — | — | not ranked |
why: cheapest competent executor; cites 3 similar verified runs. opencode has no cost evidence yet, and unknown cost is never treated as free.
Mandate mdt_51c0
- to
- agy OS-sandboxed
- scope
- src/orders/** · worktree branch only
- may
- read, write inside the worktree
- may not
- git push · network · secrets · production
- checks
- npm test -- orders
- expires
- 900 s
Execution isolated worktree
branch mandate/run_8c19e2
boundary enforced by sandbox-exec, not by the prompt
0 tokens · fresh, cache and output kept apart
executor report: PARTIAL · a report is evidence, not authority
Verification on the host
baseline, unchanged tree: 1 failing
after the mandate: 12 / 12 passing
evidence: discriminating the checks can tell the change apart
accepted · source: deterministic checks
Record
run_8c19e2 · accepted · agy · 35k tokens · 41 s · 0 human interventions
lesson stored, retrievable by similar intents
merit updated for agy in code_repo_change / medium / S
why: "agy, cheapest competent executor, 3 similar verified runs; checks discriminate."
The constitution
These articles are not settings. The system cannot amend them, and no agent can be granted the power to.
- Human sovereignty. Agents do not govern. Human intent governs; agents receive mandates.
- A fixed constitution. The rules that bind the system are not self-modifiable.
- Four different things. Capability is not competence. Competence is not permission. Permission is not a mandate.
- Least privilege. Limits are enforced outside the model, by the operating system and the control plane.
- Evidence over self-report. "Done" means verified on the host, not announced by the worker.
- Independent verification. Critical work is checked by someone other than the one who did it.
- Attribution before reputation. A failure is traced to its origin before any agent is blamed for it.
- One executor by default. A swarm is an exception that has to earn its cost.
- Honest cost. Cost is measured per accepted outcome. What is unknown stays unknown.
- Reversible improvement. Every automatic change is versioned, measured and can be rolled back.
- Local-first. Intents, runs and memory stay on your machine. SQLite is the record.
The record
Measured on one Mac, on real work, in October 2026. Small samples. We publish the limits with the numbers.
| Measure | Value | What it means |
|---|---|---|
| Agents governed on one machine | 14 | Claude Code, Codex, OpenCode and Antigravity, one agent per usable model. |
| Tests passing | 582 | Every rule above has a regression test. Live tests run separately against real agents. |
| Decision patterns registered | 52 | Each starts in shadow and is enforced only after it proves itself on measured outcomes. |
| Critical decisions served under a 50-call burst | 10 / 10 | p95 434 ms. Background calls were refused visibly, never dropped silently. |
| Tokens per accepted small edit | 4,746 → 1,435 | After economic routing, six edits per phase. Most of the gain came from ending wasteful exploration. |
| Scope-drift gate | AUC ≈ 0.5 | Measured negative. It does not work yet, so it abstains instead of guessing. |
Asked before the first mandate
Is Intentocracy another agent?
No. It does not write your code. It decides which agent should, what that agent may touch, and whether the result counts. Think of it as the government your agents work for.
Which agents does it govern?
Today: Claude Code (Haiku, Sonnet, Opus), Codex, OpenCode and Antigravity, each model registered as its own agent. Any command-line agent can join through an adapter.
Does my code leave my machine?
The control plane, its memory and its records run locally. The agents themselves still talk to their own providers, exactly as they do when you run them by hand.
How does it choose an agent?
Deterministic rules first, then verified history from your own runs, then a small local model for bounded decisions. Among competent candidates it picks the lowest expected cost per accepted outcome.
What if an agent says it is done when it is not?
Its report is evidence, not authority. The host runs the checks on an unchanged copy first and again after the change. Only checks that can tell the difference earn acceptance.
Can it change its own rules?
It can improve how it routes, always versioned, measured and reversible. It can never amend the constitution or grant itself permissions. Only you can.
Petition for access
We are opening Intentocracy to a small group of developers who run several agents every day. Tell us what you run and what you want them to do. Every petition is read.