Proving an AI agent is actually done
How do you verify that an AI coding agent actually finished the work it says it finished?
An agent reports "done" from its own account of what it did, which is not evidence. Verification means an independent check that the built thing matches the requirement: a real browser driving the feature, a test the agent did not write for itself, and a review gate each generated use case has to clear before it counts. Guardian Kane is my build of that loop, twelve numbered phases wrapped around Claude Code, and four paired A/B builds where the ungated run reported done on a doubled chart while the gated run shipped neither defect.
- Sep 21, 2026
8 min readMy Agent Shipped a Broken Chart and Told Me It Was Done
Two builds, one PRD. One rendered a clean performance chart. The other drew the same line twice and reported the task complete. How that gap turned a planning tool into a Claude Code Stop hook that makes the agent prove its work in a real browser, and what the gate caught first.
- Sep 21, 2026
8 min readThe Requirement Nobody Wrote
A four-line todo PRD said priority defaults to Medium and nothing else. One build asked how a task becomes High; the other never did. Then I built a trap for the gate and both builds walked straight past it. Four paired experiments, and the pattern they drew: the gap grows with the density of the spec.
- Sep 21, 2026
9 min readWho Verifies the Verifier?
On the densest build, eight tasks failed a browser check and one of them was a real bug. The rest were the verifier, and then my own write-up turned them into findings. Notes on checking the thing that checks, and on a three-strike retry cap the agent could quietly reset.