Proof of work for AI coding agents: a demo on every PR
Full video transcript
Your coding agent says done with the same confidence whether the fix works or not. You open the pricing page. It is not fixed. The agent was not lying: it wrote tests for its own reading of the task, and they passed. Done is a claim. Green CI is evidence only for the paths someone wrote down. A fix on the wrong card, or half a fix, passes its own tests. The cheapest evidence is a short video of the running app, attached to the pull request. A reviewer watches it in a minute, without checking out the branch. To count as proof, it comes from the running app, follows the real journey, is made after the change, and is attached to the work it proves. You should not have to record it. One line asks your agent to drive your app, narrate what is on screen, and publish it. Nothing encodes on your laptop. When the work runs as a workstream, each stage records what it did and why. The demo sits on the same card as the pull request. For a bug, record the same path twice: the bug today, then the fix. Two runs, linked by the issue URL, and one link plays both, in order. The proof lands in your Inbox, linked back to the issue it answers, so nobody has to remember which video belonged to which change. Make it the default. Add one line to your repo's agent instructions, or start the work from a recording. The full post is on the Vidmatic blog.