The autonomous engineer
What separates an autonomous engineer from a coding agent - bounded work units, an objective-to-outcome contract, and explicit limits on what is built.
A coding agent answers a prompt. An autonomous engineer owns a change from objective to outcome, under rules it cannot rewrite.
Agent versus engineer
A coding agent is a model with tools. It is as capable as its model, and its limits are whatever its host happens to enforce. An autonomous engineer in the Numinance sense is the whole system around that model: it gathers context, classifies the change, plans, hands bounded work to an agent, checks the result with independent verifiers, and records evidence. The model is replaceable. The system is what you configure and audit.
| Question | Coding agent | Autonomous engineer |
|---|---|---|
| What is the unit of work? | A conversation or session. | A change, split into work units. |
| Who decides what it may do? | The tool host, often implicitly. | A policy evaluated outside the model. |
| Who checks the result? | The same agent, or you. | Independent verifiers and a reviewer identity. |
| What remains afterwards? | A transcript. | An append-only evidence chain. |
| What happens when it fails? | It retries until stopped. | Finite repair, then escalation or rollback. |
Work units and bounded scope
The runtime does not hold one long agent session. A change is a durable state machine, and the agent implements one work unit at a time. Each work unit gets its own bounded authority: a grant naming what it may touch, at which risk class, in which environment, until when. When the unit finishes, that authority ends. The next stage starts with fresh authority, decided again by policy.
Bounded scope has three practical effects:
- A compromised or confused agent can only affect what its current grant covers.
- Parallel work units cannot widen each other's authority.
- A failure is local. It is repaired, escalated or rolled back without poisoning the rest of the change.
From objective to outcome
An objective is a request: fix this bug, upgrade this dependency, add this endpoint. An outcome is a recorded result: a pull request, a merged change, a release that was observed, or an escalation that explains why it stopped. Between them the runtime follows the change lifecycle and records each transition. The outcome is not "the agent said it was done". It is what the verifiers and evidence show.
Playbooks describe how a type of change is carried out. The bugfix-bounded playbook written by numi init is a real example: its steps are fix, review and pr-status, each bound to a different capability set, and it sets max_repair_attempts: 2. No step merges.
Designed versus available
Installing numi does not install an autonomous engineer
There are three separate things: the open numi CLI, the numinance-spec repository, and the proprietary runtime. The runtime is private and not generally available. With the open build you can validate configuration (numi validate --files) and verify evidence exports. The run, plan and approve commands are designed but absent from the open build.
See the status page for the current matrix, and Open vs proprietary for the boundary.
What it is not
- It is not unrestricted. Full autonomy means configurable, governed, qualified and human-accepted, bounded by floors, policy, grants and the breaker.
- It cannot widen its own grants, approve its own work, or promote its own autonomy level.
- It does not prove its changes are correct. Verification and evidence reduce and expose risk. They do not remove it.