Today's AI harnesses adapt people to use models for their day-to-day work — "Outlook for agents," hostage to human availability. Helix inverts it: the organisation itself is the orchestration. One loop per entity, running in the cloud, that observes, thinks, acts, verifies and learns — 24/7.
A desktop app is session-shaped: it exists while a human looks at it. The loop is event-shaped: a support email at 3am is an observation that triggers cognition without anyone opening an app.
The sidebar-agent view betrays the agent's power by queueing it behind human attention. Work should not wait for a window to be open.
Constant approval prompts optimise a large lab's PR risk, not yours. The correct optimisation is stopping irreversible actions — legal, financial, reputational — at machine speed. Everything else should just run.
Chat vs Work vs Code are a vendor's org chart, not your company's shape. Your company's shape is a loop around your policy.
The principal — founder, owner, chair — writes objectives, budgets, decision rights and prohibitions into one signed file. That's the whole interface for running the company: reach 1,000 paying customers by April, on this budget, without ever doing these things.
Policy is the only guardrail. It is enforced at dispatch — no executor, model or human, touches an actuator except through a task the dispatcher authorized against these rules.
And it fails closed: a rule the dispatcher cannot read is never an open door. Anything unparseable, unmatched or ambiguous goes to the principal.
# policies/treaty.yaml — the entity's genome objectives: - metric: paying_customers target: 1000 deadline: 2027-04-30 budgets: experiment_spend: { amount: 300, period: week, auto: true } decision_rights: - action: refund when: "amount <= 150" then: allow - action: public_claim then: gate(legal-gate) prohibitions: - impersonating_a_human_in_sales_conversations
Inboxes, analytics, uptime, bank feeds, meeting transcripts. Connectors in order of preference: API > MCP > export > computer-use > human.
Every signal lands deduped in the ledger. Nothing depends on someone noticing.
Checks each proposed task against policy and live budget counters. Decides allow, deny, approve, gate or delegate — and logs why.
Routes authorized work to the cheapest capable executor. Credentials stay in the vault; executors get scoped, short-lived capabilities.
Send the email, call the API, trigger the deploy, post the letter. The only hands the system has.
Intelligence is applied by whichever executor the task deserves. Not a hierarchy of people using tools — a router choosing among three kinds of labor.
Reflexes and routine cognition. Triage the signal, draft the reply, tag the scan. Free at the margin, so they run constantly.
Budgeted cognition and verification. The policy decides when a harder think is worth paying for.
Taste, trust, protected titles, physical and relational labor. Most organisations still prefer to buy from people. Routed by the same dispatcher, recorded in the same ledger.
An entity may touch some or all of these. The product & machinery is any owned instrument that delivers value continuously — software, a car-wash machine, plant machinery. A pure service company's instruments are mostly humans. Physical inventory is state in the world, not data; buying stock is capital allocation under policy counters.
Most systems observe → think → act. An effective organisation also verifies and recursively improves itself. Verification is an independent identity: the actor can't pass its own work, the verifier can't fix it. What verification learns becomes a lesson, proposed as a PR against the policy repo — reviewed and signed by the principal, or auto-approved where the principal has configured that per entity.
"The limitation in the loop is going to be gathering evidence, rather than cognition speed."
Entities are still owned by humans — the principal bootstraps the policy, signs the lessons, and holds every gate. Everything else runs.
Sign in