I Built a Multi-Agent Coding Setup to Keep Me in the Loop
Why I built this I didn't build this after a big disaster. No agent ever deleted my database or broke a production deploy. The problem was slower and more annoying than that. Every time I used an AI agent on a real task, I spent too much time managing it. It asked permission for simple things, the implementation often failed when the task was bigger, and even when the code worked, it often…
In response to the frustrations he experienced with AI agents managing his workflow, the author created Jancera/skills, a suite of skills and subagents designed to keep them in the loop during coding tasks. These agents aim to reduce the tedium of managing AI agents, which often asked permission for simple tasks, failed to implement larger projects correctly, and frequently ignored the project's organization.
Three main frustrations led to the creation of specific rules within the system: 1) excessive permission prompts, 2) challenges with implementing large tasks, and 3) agents never merging changes.
Rule #1 addressed the numerous permission prompts the agent received, with read-only inspection commands now approved by default. Implementations now run within an OS-level sandbox, where commands execute without prompts and network access is only permitted when necessary, such as for commits, git pushes, package installations, file downloads, or external edits.
Rule #2 tackled the issue of large tasks being poorly implemented. By breaking features into phases and further dividing them into vertical tasks, the author ensures each task touches only the necessary components and produces a runnable, testable result. These tasks run in separate git worktrees from the same base branch, preventing overlapping work.
Rule #3 prevented agents from automatically merging changes. Instead, the author reviewed the combined diff of each phase for bugs, security issues, and regressions. Once all tasks in a phase were merged, an oracle subagent checked the phase for any problems, forwarding fixes to a fixer agent or UI work to a designer agent.
To test the effectiveness of the setup, the author applied these rules to a project called shorts-pipeline. They initiated four tasks simultaneously—subtile settings in the web UI, target shorts per project, regenerating audio for a single video, and a configurable desired video length. These tasks, totaling about 1,360 lines of code, were completed in roughly a day.
The author did not reject any of the merged changes and pushed them to the main branch, after which the oracle reviewed the combined diff and discovered critical issues that none of the individual tasks caught. These findings led to further improvements before the code was finalized.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
