Kata Agent Team
Autonomous coding agents that continuously improve
An agent team that improves itself on a daily Plan-Do-Study-Act cycle. It ships eighteen skills and a small agent roster, and you maintain no infrastructure.
Agent teams fail when they get complicated.
Most agent setups grow heavy infrastructure, large toolchains, and prompt chains that no one can audit. Kata keeps the setup small, so you get the parts the team needs and nothing else.
The team runs one cycle every day.
Each workflow belongs to one phase of the cycle. Every finding from the Study phase goes back into the loop as an action.
Turn an approved spec into a design, and then turn the design into a plan with steps, files, sequence, and risks.
Carry out each plan through implementation pull requests, and run the scheduled workflows that harden, release, and maintain the repository. Every run records a trace.
Read the output in four streams: security audits, feedback triage, documentation review, and grounded-theory analysis of traces.
A small finding becomes a fix pull request, and a structural finding becomes a spec document. The two kinds of branch stay separate.
Each of the eight agents has a written scope.
Each agent profile states what the agent does and what it must leave alone. When an agent finds a problem outside its scope, it writes a spec instead of a fix.
Takes an approved spec through design, plan, and implementation.
Keeps pull request branches ready to merge, repairs CI, and cuts releases. It is the only agent that merges outside contributions.
Patches dependencies, secures the supply chain, and applies the security policies.
Audits code health, reviews maintainability, and removes debt without a change in behavior.
Triages issues against the product vision, reviews spec quality, and runs evaluations.
Reviews the docs for accuracy, curates agent memory, and fixes stale pages.
Retires old logs, storyboards, and completed or cancelled specs after their useful content is saved elsewhere.
Runs the daily storyboard meeting and the one-on-one coaching sessions.
The same agents work on every surface.
Your IDE, a cron schedule, a GitHub event, or a bridged chat message all start the same profiles and skills, and the agents behave the same way in each case.
Shared memory is one Git repository of markdown files.
Every agent reads and writes the same wiki, which holds priorities, logs, metrics, and storyboards. A scheduled shift, a message from a bridge, and an IDE session all share this state. The wiki needs no database, because it is a set of markdown files in a Git repository.
Set up the team with three commands.
Install the skill pack, then ask Claude to set up the team.
Follow the full path in Get started, then read the documentation.