Member-only story
I rebuilt my agentic coding kit again because I had made the exact mistake I keep seeing in coding-agent setups: I had built enough process to make small work feel expensive.
A one-file correction could become a scout, a plan, a coder, several reviewers, a memory update, a reflection, and another round of handoffs. Each step sounded sensible in isolation. Together, they created a context bill and a coordination bill that the task never earned.
The new version, Agentic Coding Kit v6, has a much leaner rule: the main coding session owns the outcome. Bounded agents answer narrow questions. The loop expands only when the evidence says it should.
I wrote earlier about loop engineering. The core argument was that agent definitions could stay thin while loop bodies carried the contract, verification, repair, budget, and stopping logic. I also reported a 120 out of 120 versus 112 out of 120 result for my kit against OpenCode pure mode on a 24-task executable suite. The margin was useful, but it was smaller than I expected.
That earlier rebuild taught me where the value lived. This one is the correction: even a loop-first kit becomes wasteful when every task enters a deep loop, every specialist receives too much context, and temporary workflow material…