Kata Agent Team
Autonomous coding agents that continuously improve
An autonomous agent team that keeps getting better — organized as a daily Plan-Do-Study-Act cycle. Eighteen skills. A focused agent roster. Zero infrastructure.
Agent teams fail when they get complicated.
Most agent setups drown in infrastructure, sprawling toolchains, and unauditable prompt chains. Kata takes the opposite path — radical simplicity. Everything you need, nothing you don't.
One cycle. Every day.
Every workflow maps to a phase. Findings from Study always re-enter the loop — nothing is observed without downstream action.
Turn approved specs into architectural designs, then executable plans with steps, files, sequencing, and risks.
Execute plans via implementation PRs. Run scheduled workflows that harden, release, and maintain. Every run captures a trace.
Analyze outputs across four streams: security audits, feedback triage, documentation review, and grounded theory from traces.
Trivial findings become fix PRs. Structural findings become spec documents. Fix and spec branches never mix.
Eight agents. Explicit scope.
Each persona knows what it must do — and what it must not. When a finding exceeds scope, the agent writes a spec rather than attempting the fix.
Owns the full spec, design, plan, implement arc for approved specs.
Keeps PR branches merge-ready, repairs CI, cuts releases. The sole external merge point.
Patches dependencies, hardens supply chain, enforces security policies.
Audits codebase health, reviews maintainability, and clears debt without changing behavior.
Triages issues against product vision, reviews spec quality, runs evaluations.
Reviews docs for accuracy, curates agent memory, fixes staleness.
Retires stale logs, storyboards, and completed or cancelled specs once their signal is safely preserved.
Facilitates the daily storyboard meeting and 1-on-1 coaching sessions.
Same agents. Every surface.
The same profiles and skills operate identically whether triggered by your IDE, a cron schedule, a GitHub event, or a bridged message.
One Git repo of markdown files.
Every agent reads and writes the same wiki — priorities, logs, metrics, and storyboards. A scheduled shift, a bridge-dispatched message, and an IDE session all share the same state. No database. Just markdown in a Git repository.
Three lines. That's it.
Install the skill pack. Tell Claude to set up the team.