Six Months of Writing Code Exclusively With Agents

In February Maisem Ali made himself a rule: no more writing code by hand. Six months on, this is the account of what that forced him to build. The discipline is the interesting part. When an agent got stuck, finishing the job himself was not allowed, so the only move left was to improve the prompt, the tools or the environment until the agent could get there alone.

That constraint scaled him from one agent to about twenty concurrent VMs, each task in its own isolated machine, which in turn needed botd, a control system reachable from his phone that provisions the machines and keeps every conversation.

The claim people will argue with is that peer code review is dead at exe.dev. They do not review diffs; they merge what they think should be merged, on the grounds that the review that matters already happened in the design discussion, the contracts and the validation. Agents do a second pass, but the human judgment is about whether a change belongs at all, not whether the lines are right.

The supporting practices are more quietly useful: design before implementation, behaviour-driven tests instead of tests that mirror the implementation, migrations treated as the main risk, and patterns enforced by custom tooling rather than convention. He is candid about the cost, saying he has lost the granular familiarity with his systems and now queries agents to get it back. Read next to Duckbill’s risk-based version of the same idea, which keeps humans on the changes that can actually hurt.