Experiment · Agentic engineering
EngTeam Wallkit & Blueprint
An open, dependency-free scaffold that stands up a complete software engineering organization operated by LLM agents — architecture, product and project management, analysis, development, QA, release engineering, security, compliance and metrics — as one written, auditable, closed-loop process.
This is a kit and blueprint for a complete engineering team run by AI — and that is precisely why this notice exists. The author does not endorse using it to replace human decision makers, builders, architects, designers, project or program managers, engineering leaders, or any other role — specifically or arbitrarily. Every gate in this kit that routes to “the engineer” exists because a human owns that call, and removing the human removes the safety property, not just the person.
The recommendation: review this kit and blueprint with your existing team, and use these tools to help that team deliver and support quality, secure, scalable, enterprise-grade products — an amplifier for the people accountable for the work, not a substitute for them.
The experiment, stated honestly
The framing is not “replace your engineering team with AI” — it is the opposite question, asked rigorously: where, exactly, are the gaps that need humans?
The kit deliberately builds the entire org chart in agent form, because you cannot find the load-bearing human decisions until you have tried to write down everything else. Every place the process routes to “the engineer” is a finding: a decision class the experiment concluded no agent may make. The kit, the blueprint and the current answers are public.
What is underneath the wall
- A role roster with written authority boundaries — Maestro, Architect, Builder, Reviewer, Researcher, Warden, Foreman, Adjudicator, and the human Patron.
- An append-only, byte-reproducible event ledger with integrity checking: status is counted from the repository, CI and the harness, never asserted by an agent about itself.
- Dispatch and handoff procedures — leases, and a single serialized PR slot.
- A decision log (DEC-NNNN) that survives any individual session.
- Traceability from requirement to change to test to release, answerable by command.
- A security lane in the definition of done — SAST and secrets scanning with per-rule promotion, secrets requiring rotation rather than deletion, researcher network access off by default, the wall server bound to loopback, no tokens baked into images.
- A compliance register — SOC 2, HIPAA/PHI, PCI, PII/privacy, government and sector — with self-attestation, and a Warden that can block autonomously but can never grant new access.
- Deployment targets written for bare hardware, VMs, Docker and Kubernetes.
Nine tabs, and every number counted
The visible surface is a live status wall: MAIN, STORIES, CREW, LEDGER, WAITING, RETRO, POSTURE, DOCS and FLOW, plus the compliance self-attestation and iteration drill-in popouts. Every screen on the projects page is a real capture; the reference deployment’s board carries 1,143 items.
Read the kit’s guide, in its README on GitHub →