PGAI AGENT KANBAN

An autonomous task-decomposition framework
that ships its own releases

public release: v2.4.0 · MIT

Drop a requirements doc. Walk away. Come back to a tagged release.
Specialized LLM agents decompose, implement, verify, tag, and push — unattended, on your own VPS, on the LLM provider of your choice.

The Thesis

RELIABLE

  • predictable
  • reviewable
  • slow

AUTONOMOUS

  • unsupervised
  • fast
  • surprising

Decompose work into small verifiable tasks. Specialize agents per role. Let files on disk be the truth and git the safety net. Single-thread by design. An 85% solution that keeps moving beats a 100% solution that stalls.

Six Agents + an Overseer

POturns a plain-language brief into a formal requirements doc
PMdecomposes requirements into tasks
CODERimplements in an isolated git worktree
WRITERdrafts prose deliverables and release notes
TESTERverifies against the requirements — and files its own bug reports
CMbranches, tags, releases — the only agent that touches origin
OVERWATCHhourly zero-token health sweeps, a daily LLM deep-clean, and a trigger that wakes it whenever any agent blocks
requirements ─▶ decompose ─▶ rc/vX.Y.Z ─▶ build ─▶ verify ─▶ squash ─▶ tag ─▶ vX.Y.Z live on origin

Three Workflows — or Author Your Own

RELEASE

Software with the full git lifecycle: RC branches, gated test suites, tagged releases, generated release notes and CHANGELOG.

DOCUMENT

Prose deliverables — reports, stories, specs — published as versioned artifacts with the same metrics and audit trail.

TESTING-ONLY

Audits something already shipped: a read-only verification of any project's tag. Produces a report, never a release — and files genuine findings as bugs on the target's own lane.

Workflow types are plugins — a directory with a manifest and eight hooks. A generator scaffolds new ones; the engine asks capabilities, never names. The tutorial builds a throwaway type in ~15 minutes.

Proof, Not Promises

IT BUILDS ITSELF

You're looking at v1.23.0 — shipped by the versions before it. Well over a hundred releases of this codebase were decomposed, implemented, verified, tagged, and pushed by the framework, autonomously. Self-build isn't a mode: the framework's repo is registered like any other project.

IT DEBUGS ITSELF

In launch week alone, 60+ bugs were found, filed, fixed, and shipped through the pipeline — many filed autonomously by TESTER against the framework's own code. The gated suites now run 2,139 unit and 191 integration tests plus ten structural lints on every release.

IT AUDITS ITSELF

The testing-only workflow was certified end-to-end — selection to report to closure — with zero operator interventions, verifying another project's shipped tag and filing real bugs it found there. The audit is ephemeral; the findings are not.

Run It

# your VPS, your API key, your choice of LLM
$ git clone https://github.com/purdygood-ai/pgai-agent-kanban.git
$ cd pgai-agent-kanban && ./install.sh
# then walk one of the three included demos:
$ less demos/chomp-man-demo/README.md

Three demos ship in the repo — build a game, write a story, then audit the game you built. Everything is open: docs, role files, workflow plugins, the tutorial, all demo inputs.
Your hardware. Your data. Your LLM.

The Economics

~$200 / yearVPS — hardware is solved
~$100 a dayLLM tokens (or use oauth)
~$200 a monthflat-rate OAuth plans

Shipped Since v1.0

The Road Ahead