Agent-Legible Software¶
Ryan Lopopolo's operating principle during the 0%-human-code experiment: software must be written, structured, and documented so that a coding agent can navigate and modify it, not (primarily) so that a human reviewer can. A natural extension of harness-engineering — if agents do the writing and reviewing, they are the audience.
Concrete moves Ryan describes:
- Test-the-source: keep files ≤350 lines so a whole file fits cleanly in a context window
- Bespoke lint rules encoding codebase-specific invariants (e.g. "fetch without retry/timeout is an error") — the human-taste layer compiled into something an agent can check
- Observability built for agents, not humans: traces, metrics, and logs the agent can actually read and act on, not just a dashboard
- "Don't accept slop" — absorb the short-term velocity hit to install durable guardrails
Contrast with legacy "human-legible" practices (DRY, clever abstractions, reviewer-friendly PRs) which can actively harm agent throughput. Pairs with non-functional-requirements-as-prompts — if agents write code that satisfies explicit NFRs, the NFRs must be legible to them first.