Engineering Playbook
EXECUTE
Delegate according to risk. Engineering discipline remains mandatory even when AI increases velocity.
Speed without structure creates debt. Execution should be guided by specs, tests, clear ownership and explicit limits on agent authority.
What breaks without it
Without structured execution, speed creates debt: uneven quality, weak ownership and systems that cannot be operated safely.
The deliverable
The Autonomy Model
The question is no longer whether AI can write code. It is how much authority an agent should receive for this specific change.
Engineering is shifting from AI-assisted development — "write this function" — toward delegated agentic engineering: take this issue, understand the repository, implement it, run the tests, fix the failures and open the change.
That shift makes a single question load-bearing: what is this agent allowed to do without asking? Answering it once, as policy tiered by risk, is what makes delegation safe enough to scale. Answering it ad hoc, per prompt, is what makes it dangerous.
Autonomy should be a property of the change, not of the tool. A documentation fix and a payments migration deserve different answers even from the same agent.
- A0 — Read only: investigation and explanation, no write access
- A1 — Autonomous: documentation and trivial change, standard review at merge
- A2 — Autonomous with mandatory verification: normal change, evidence pack required
- A3 — Human approval before execution: auth, payments, infrastructure, migrations
- A4 — Human-controlled execution: production, destructive and regulated actions
Autonomy levels
How much authority for this change?
Autonomy is a property of the change, not of the tool. A documentation fix and a payments migration deserve different answers from the same agent.
| Level | Examples | Agent autonomy | Control |
|---|---|---|---|
| A0 Read only | Explanation, investigation, code search, impact analysis | Agent may read and reason. It may not modify anything. | No approval required. No write access granted. |
| A1 Autonomous | Documentation, comments, formatting, test-only additions, trivial dependency bumps | Agent implements and opens a change without prior approval. | Standard review applies at merge. Automated checks must pass. |
| A2 Autonomous with mandatory verification | Normal application change, refactor within an owned boundary, bug fix | Agent implements, but the change cannot merge without a complete evidence pack. | Tests, static analysis and security checks are non-negotiable gates. Human reviews evidence, not only diff. |
| A3 Human approval before execution | Authentication, authorisation, payments, data migration, infrastructure, cross-boundary changes | Agent may plan and propose. It may not execute until a human approves the plan. | Approval is on the execution contract, not on each action. Named human accountable for the decision. |
| A4 Human-controlled execution | Production operations, destructive actions, regulated changes, secrets and key material | Agent may prepare and document. A human executes. | Explicit human action required. Fully audited. No delegated credentials. |
Practices
- Grant least privilege by default and widen deliberately
- Make bypass paths exist, but auditable — an unusable rule gets routed around
- Enforce standards with deterministic hooks where judgement is not required
- Keep changes small enough that a reviewer can hold the whole diff in mind