Spec-driven development: how we review plans, not pull requests
Dominik Schmidle · 13 October 2026
Read the originalGiant Swarm moved human review from code changes to a written plan, so people settle intent before AI agents build and most code merges unread.
Giant Swarm helps platform teams run Kubernetes, software for running applications across many servers. Pull requests from its staff and their AI agents tripled in a year, so the team now reviews and merges a written plan before agents implement it. Dominik Schmidle describes the steps, and the company has released the workflow as open source.
- Monthly pull requests rose from about 700 in summer 2025 to more than 2,100 in summer 2026, not counting automated bots.
- An AI agent checks each factual claim in the plan against the real code; in the published example, one of eight claims named a configuration setting that did not exist.
- A separate step argues for building less, and most plans come back with instructions to simplify.
- Most code changes merge without a person reading them, while changes to public interfaces or security still get human review.
A team that cannot keep up with agent-written code changes can review the plan and the tests instead, provided agents can run those tests in a realistic local environment. The piece is a company's account of its own process.
Review and merge a written plan for any change that contains a decision someone could disagree with, then let AI agents implement against it.
Have an AI agent check every factual claim in a plan against the real code before people approve the plan.
Derived by Working Surface from the article; more in the Playbook. Source line: We moved human review from the code diff to the plan for what we're about to build.
15 October 2026: when AI agents make building cheap, the decision worth reviewing is the one made before the build, written where other people can dispute it.