Documentation

The platform runs one loop. You stand up the environment, run sessions, look at what happened, keep a record of it, measure the outcome, and stop the team when it produces more than a human can read. Each section below covers one part of that loop. Start with the first one, then choose the guide for the task you have now.

Stand Up the Platform (Teams Using Agents)

Getting Started: Your First Trace

Install the Gemba skill pack and the gemba-* command family, run one agent session, and read the NDJSON trace it captured.

Coordinate an Agent Team (Teams Using Agents)

Coordinate an Agent Team

Run a lead and N participant agents in one asynchronous session. Choose supervise, facilitate, or discuss. Pass messages with Ask, Answer, and Announce. One NDJSON trace records everything that happened.

Prove Agent Changes (Teams Using Agents)

Prove Agent Changes

Reproducible evidence that agent changes improved outcomes, from the eval session through the trace analysis.

Run an Eval

Run an agent-as-judge eval in CI and get a traceable verdict on whether an agent change improved outcomes.

Run a Benchmark

Prove a skill-pack change improved coding outcomes. Run a task family across N runs, grade with hidden tests, and report pass@k.

Automate with GitHub Actions

Run gemba-benchmark in CI with the forwardimpact/gemba-benchmark composite action. You get step summaries, artifact upload, and PR-triggered benchmarks.

Analyze Traces

See exactly what an agent did and why. Download traces, query turns, filter by tool or error, and measure token cost.

Operate a Predictable Agent Team (Teams Using Agents)

Set Up Persistent Memory and Metrics

Give your agent team persistent memory and real signal detection with wiki-backed state and XmR control charts. The team then acts on real changes instead of noise.

Send a Memo or Update a Storyboard

Communicate across your agent team and keep storyboards current. You do not manage the wiki infrastructure yourself.

Audit and Auto-Fix the Wiki

Keep the wiki valid against a declarative rule catalogue. Auto-fix what is safe to fix. Report the rest to a human, so that stale memory does not corrupt coordination.

Allocate Collision-Ledger Entries for Parallel Work

Assign stable ids to parallel work without merge collisions. An append-only issue thread anchors every id. The ledger page gets a projection only when you rebuild.

Chart a Metric and Check Variation

Know whether a metric changed or only varied. Natural process limits and Wheeler's detection rules separate signal from noise.

Guard an Agent Team (Teams Using Agents)

Guard an Agent Team's Activity

Bound an agent team's output volume with a deterministic brake. Four counters over one window set an operator latch that only a human clears.

The platform ships commands and CI actions. It has no importable API of its own. When you want the components instead of the commands, import the runtime libraries directly. The library catalog lists that API surface.

Kata is the reference tenant, and it runs this platform every day. The Kata site documents the agent-team practice that uses the loop.