Our shared agent workflow now includes review and repair

DocEngine now runs a shared agent workflow on a self-hosted OpenHands environment, where agents implement GitHub Issues tasks, open pull requests, and run a review-and-fix loop before a human decides whether to merge. Merging triggers automatic deployment, and the team is evaluating the no-mistakes tool, whose auto-fix loop has attempt limits and pauses for human decisions, as a possible addition. Approval to merge remains with the humans, the author wrote, making repeated review and repair a working part of the self-correcting autonomous delivery stage outlined in an April 2026 post.

We now run an OpenHands https://github.com/OpenHands/OpenHands -based agent environment in our own cloud. We have one place to change the context our agents use, their skills, and their plugins. In April, I wrote about moving from individual AI assistants to a shared delivery workflow https://hajek.no/posts/2026/from-ai-assistants-to-agentic-delivery-workflows . Our target was an issue-driven process where agents could implement changes and prepare pull requests for human review. Presentation slide for the five-stage framework discussed in my April 2026 post. At the time, DocEngine was at Level 2, with Level 3 as the next target. That process is now running. We control the agent environment and can configure agents with different specializations. GitHub Issues is our primary entry point, although we can also start in the OpenHands interface. Before implementation, we can discuss the task with the agent and agree on what it should deliver. From there, the workflow is: 1. The agent implements the agreed task and opens a pull request. 2. It reviews the change, fixes the issues it finds, and runs that review-and-fix loop again. 3. A human reviews the resulting pull request and decides whether to merge. 4. Merging triggers automatic deployment. We can also add tools to this environment. no-mistakes https://github.com/kunchenguid/no-mistakes , for example, can run validation, ask an agent to fix findings, and check again. Its auto-fix loop https://kunchenguid.github.io/no-mistakes/concepts/auto-fix/ has attempt limits and pauses for human decisions. It is a possible addition to our existing workflow. The final stage I described in April was self-correcting autonomous delivery loops. Repeated review and repair gives us a working part of that stage. Approval to merge remains with us.