Code Review as a Principal-Agent Problem

Code review (in the review-then-commit sense — Apache 1990s, Google 2000s, GitHub PR 2010s) is a low-trust collaboration mechanism: a contributor proposes a change, a reviewer LGTMs it, the change lands. It works for teams beyond the two-pizza size and for OSS for a single reason: the reviewer can cheaply infer the contributor's effort by reading the diff. The diff is testimony.

When the contributor is an agent driven by a human typing two sentences, that signal collapses. The reviewer can no longer tell whether the contributor:

• Understood the constraints
• Considered alternatives
• Tested edge cases
• Spent thirty seconds or three days

This is the standard principal-agent problem (Wikipedia) — the principal (reviewer) can't observe the agent's (contributor's) effort, and effort costs fall on the principal. Before agents the imbalance was bounded; after agents it's unbounded.

The visible failure modes:

• OSS slop PRs — agents generate plausible PRs at zero contributor cost; maintainers burn out responding
• Internal review queues — companies that mandate review can't extract agent productivity gains because review bandwidth becomes the binding constraint
• Reviewer-as-prompter — reviewers realize they could re-prompt the agent themselves faster than reading the contributor's output

Process responses crawshaw-agent-principal-agent catalogs:

1. Small high-trust team, no second reviewer — human-prompts-agent + human-reviews + human-deploys. Works because the principal is the agent. Requires investing in integration/e2e/agent-driven test machinery.
2. Pre-2005 Microsoft model — independent teams, QA-driven sync, no mandated review-before-commit. Worked at scale but low-trust BigCos won't reach for it.
3. The OSS retrofit — replace "send PR" with "send prompt + feedback loop"; see i-dont-want-your-prs for Ciężarkiewicz's version.

The unresolved problem is large-company low-trust environments. There is no known process that gets review out of the bottleneck position there; reviewer bandwidth ceilings + doubled review = agents producing little net productivity gain. This is part of no-silver-bullet-llms's thesis from a different angle.

Adjacent: contributor-poker (Cro's frame: bet on the contributor not the contents of the PR — itself a response to the same effort-inference collapse), skill-atrophy-supervision-paradox (the supervisor-side problem).