The agent principal-agent problem
a day ago
- #AI-agents
- #code-review
- #software-development
- Traditional code review processes like review-then-commit were effective for low-trust collaboration among engineers, popularized by GitHub PRs.
- Agents disrupt this by doubling review workload and increasing changes, leading to bandwidth shortages and the principal-agent problem, where reviewers can't infer effort.
- Small high-trust teams can adapt by having humans own deployment, using tests and agent workflows, but large companies struggle due to low trust and blame-avoidance.
- No scalable process for agent-driven development in large companies exists yet, though historical examples like 1990s Microsoft show alternatives can work.
- Small teams gain a force multiplier from agents, while large teams face productivity drains; adoption is limited without new robust processes.