Documentation
Kata is a practice for a team of coding agents. The team works one cycle a day, and each job below covers one part of that cycle. Start with the setup job. Then pick the job you have in front of you now.
Set Up the Team (Teams Using Agents)
Run a Continuously Improving Agent Team (Teams Using Agents)
Run a Continuously Improving Agent Team
Run the daily Plan-Do-Study-Act cycle with an agent team. Every shift leaves evidence behind. Every finding re-enters the loop as a pushed fix or a pushed spec.
Choose and Scope Your Agent Roster
Decide which agent personas run in your repository. Write the scope constraints that turn an out-of-scope finding into a spec instead of an unreviewed fix.
Keep Team Memory and Coordination Apart
Memory holds an agent's own state. Coordination needs a named receiver and an addressable artifact. Learn where the line sits. Learn what breaks when a team buries a handoff in the wiki.
Run the Daily Storyboard and Coaching Session
Facilitate the team storyboard meeting and the one-on-one coaching session with the Toyota Kata five questions. Every experiment then starts from measured data instead of an impression.
Turn Study Findings into Fixes and Specs
Classify every Study finding once. Then route it to a fix branch, a spec branch, a Discussion, or a close. Fix branches and spec branches never mix. A local commit is never a completion signal.
Take a Change from Spec to Shipped (Teams Using Agents)
Take a Change from Spec to Shipped
Walk one change through the full arc: spec, design, plan, implementation, independent review, merge, and release. Each phase produces one artifact that answers one question.
Set the Approval Gates and Trust Boundary
Decide what your agents may approve and what a human must approve. Then make the approval signal travel from a human action to the state file the agents read at merge time.
Kata ships skills and agent profiles. It ships no command of its own. Three other standards carry the parts underneath it, and each one has its own site.
-
The runtime is Gemba.
Kata skills call the
gemba-*commands to run a session, read a trace, hold memory, and chart a metric. Read every command surface there. - The instruction architecture is Jidoka. Use it when you author or repair an agent profile, a skill file, or a checklist.
- The repository shape is Monorepo. It names the root files and the directory tree an adopting repository carries.