What we actually let AI agents do (and what we never will)

We are frequently asked about the roles of agents and humans in our work, often with skepticism: "Do agents really write less than you claim?" or "Do they write more than you admit?" To clarify, here is the policy from our delivery handbook — this outlines the actual rules our engineers follow, along with the reasoning behind them.

01Agents handle the mechanical layer

• —Scaffolding and project structure: Generated by agents, then reviewed once by a human before any further development occurs.
• —Test suites: Agents write the majority of unit and integration tests, including tedious edge cases that humans often overlook.
• —Boilerplate code, data transfer objects, API clients, and migrations: Fully generated by agents and reviewed by humans through a diff process.
• —Documentation drafts: Agents create initial drafts, which are then corrected by engineers — this correction is where the documentation becomes accurate.
• —First-Pass Code Review: Agents can summarize an unfamiliar codebase or pull request more quickly than any human can skim through it.

02Humans control the judgment layer — always

• —Architecture and data modeling: Agents may propose designs, but the final design that is implemented is authored by a human who can justify every decision.
• —Security-Relevant Changes: Matters involving authentication, cryptography, secrets handling, and permissions are either human-written or heavily rewritten by humans, period.
• —Anything touching production data or customer systems: A human is responsible for executing or approving changes, with a full audit trail in place.
• —Trade-Off Decisions: Decisions regarding scope reduction, acceptance of technical debt, and defining what "done" means for a client — that's precisely why clients hire us, not our tools.

03The line we will not cross

No change will be made to a client repository without a human reviewer who comprehensively understands it and is accountable for it. This policy is not due to any shortcomings of agents — rather, accountability is our product. A client investing in production software expects someone to take responsibility when issues arise, even at 2 AM. That responsibility rests with a person, and our process ensures that individual can always truthfully say, "Yes, I understand this change."

04Why this is a durable policy, not a transitional one

Every quarter, agents improve, and the mechanical layer becomes more sophisticated. However, the judgment layer does not diminish in the same way — it becomes more leveraged. The exact boundary of our policy may shift, but its fundamental shape remains constant. The future of this field is not about humans versus agents; it's about providing a clear and defensible answer to the question, "What do you still insist a human do?" We insist on retaining human involvement in areas where accountability for the work is paramount.