Kakuna Automates Codebase Hardening with Parallel Subagents
Agentic tool Kakuna tackles codebase maintenance autonomously, returning a self-audit after a day of parallel subagent work.
swyx shared details of Kakuna, an agentic coding tool built around a two-phase workflow for automated codebase hardening. A developer runs /plan to scope the work, then triggers /goal and steps away; after running autonomously for roughly a day, Kakuna returns with feature functionality preserved, maintenance tasks completed, and a self-generated audit of everything it changed.
The tool is designed around subagent parallelism — multiple agents working concurrently rather than a single agent running sequentially — and is structured as skills paired with checklists. The self-audit output is a notable design choice, giving developers a traceable record of changes rather than an opaque diff. For teams carrying a backlog of non-functional work — the hardening, cleanup, and hygiene tasks that accumulate in any active codebase — Kakuna represents a pattern where that category of work is delegated wholesale to an autonomous agent, with the developer reviewing results rather than doing the work.