What they did
Kento Ikeda, who runs BrightFlow LLC, built a system he calls BRIDGE to develop his company's product with two AIs. ChatGPT, connected through a custom MCP app, gathered context, wrote requests and reviewed work, but held no GitHub token or Kiro key. It could call only five tools: two to inspect the repo and PRs, one to hand a task to Kiro, one to check a task, and one to merge. Each task had to declare which files Kiro could create, change or delete. A controller then ran Kiro CLI headless inside a bubblewrap and seccomp sandbox with no shell access, compared the real diff against the declared list, failed the task on any undeclared change, and opened a PR. Merges went through only if the PR's head commit matched the one that was reviewed. Ikeda set tool confirmations to always allow and stepped in only for direction or when the AI got stuck. He ran it from 11 September to 12 October 2026.
How it went
The month produced 317 merged PRs, 159 closed and 148 still open. Of the 307 unmerged, 122 had no diff at all, because ChatGPT used the write path for audits and questions when the read tools could not answer them, and 37 were superseded by later PRs. The median time from PR to merge was 32 seconds and there were no GitHub reviews; merges rested on the AI's review. 19 PRs merged after main had moved on without being rechecked, though none overlapped later changes. Ikeda reports no write ever escaped the declared scope or the sandbox. The controller grew to about 49,000 lines of TypeScript, half of it tests.
Worth knowing
Ikeda logged so little, to protect secrets, that he cannot count how often the scope and SHA checks actually stopped something; record result types and counts from day one. Give the planning AI a read path for questions or it will open empty PRs. PR diffs shown to the reviewer were capped at 128 KiB, so the AI may not have seen whole changes. This is his own setup, not one AWS recommends.