What we actually let AI agents do (and what we never will)
"AI-assisted development" is uselessly vague. Here is our actual autonomy policy: the exact lines we draw, and why they sit where they do.
We get asked this constantly, usually in the form of a suspicion: "surely the agents write more than you admit" or "surely they write less than you claim". So here is the actual policy from our delivery handbook — not marketing, but the rules our engineers work under, with the reasoning attached.
01Agents own the mechanical layer
- —Scaffolding and project structure: generated, then reviewed once by a human before anything builds on it.
- —Test suites: agents write the bulk of unit and integration tests, including the tedious edge-case matrix humans skip.
- —Boilerplate, data transfer objects, API clients, and migrations: fully agent-generated, human-diff-reviewed.
- —Documentation drafts: agents write, engineers correct — the correction step is where the doc gets honest.
- —First-pass code reading: agents summarize an unfamiliar codebase or PR faster than any human can skim it.
02Humans own the judgment layer — always
- —Architecture and data modeling: agents may propose, but the design that ships has a human author who can defend every choice.
- —Security-relevant changes: auth, crypto, secrets handling, and permissions are human-written or heavily human-rewritten, full stop.
- —Anything touching production data or customer systems: a human executes or approves, with an audit trail.
- —Trade-off decisions: when to cut scope, when to take on tech debt, what "done" means for this client — these are why clients hire us, not our tools.
03The line we will not cross
No change reaches a client repository without a human reviewer who understands it well enough to be accountable for it. Not because agents are bad — but because accountability is the product. A client paying for production software is paying for someone to answer for it when things break at 2am. That someone is a person, and our process is built so the person can always honestly say "yes, I understand this change".
04Why this is a durable policy, not a transitional one
Every quarter, agents get better and the mechanical layer grows. The judgment layer does not shrink the same way — it gets more leveraged. Our policy's exact boundary moves; its shape does not. The future of this discipline is not humans-versus-agents but a crisp, defensible answer to "what do you still insist a human does?" We insist on the parts where being able to answer for the work is the work.