Open source · CC BY 4.0

A governed operating model for AI-assisted delivery.

Durable context, autonomous worker sessions, explicit human decision rights, isolated execution, quality gates, auditable provenance. Mother CLAUDE runs a team of AI coding sessions that ship real work — while a human keeps every decision that genuinely needs one.

4.2×
tickets / active day vs. baseline
12+
repositories governed
4
roles: worker · dispatcher · steward · human

01 — The system in one picture

The tracker is the coordination layer. Nothing else is.

HUMAN
product · UX · risk · prod promotion
DISPATCHER
specs features → Ready tickets
ISSUE TRACKER
durable state + audit trail
worker
worker
worker
each in an isolated git worktree
PR + CI
STEWARD
review · reconcile · merge or escalate
01Dispatcher feeds the queue

A human and one interactive session spec features into self-contained Ready tickets: context, scope, acceptance tests, pointers.

02Workers drain it

Claim → reconcile → build. Each worker runs in its own git worktree on a per-ticket branch, so sessions never collide.

03CI is the evidence

Every ticket lands as a PR into the integration branch. Tests and CI decide whether work is objectively done.

04Steward moves it through

Green and objective merges automatically. Anything touching UX or judgment is escalated to the human with verify steps.

05The human decides and verifies

Product direction, visual judgment, production promotion. Chat is for attention only — never for state.

To DoReadyIn ProgressNeeds InputIn ReviewChanges RequestedDone

02 — Three layers

Memory, work, governance — stacked in that order.

Layer 1

Memory

A three-tier documentation architecture: lean per-project context, shared cross-project standards, deep docs loaded on demand, plus session handoffs that survive an ephemeral session.

Layer 2

Work

The issue tracker as the only coordination layer — roles, queues, claims, isolated worktrees, PRs. No session talks to another; status is the single source of truth.

Layer 3

Governance

Definition of Done, tests and CI as objective evidence, steward review, escalation paths, graduated trust, and explicit human decision boundaries.

03 — Governance

What the AI is not allowed to decide.

Autonomous sessions may make reversible engineering decisions, constrained by documented standards and tests. Everything else escalates. If it takes taste to judge, a human judges it.

“LLMs don't need more prompts — they need better institutional memory. Autonomy needs governance before it needs scale.”
Product direction and UX judgment
always human
Production promotion
human-gated after verification — agents never merge to production
Cross-system architecture decisions
shared blast radius
Dependency and infrastructure changes
a shared dependency tree is shared mutable state
Security and data-governance exceptions
no autonomous exceptions
Anything without objective evidence
if it takes taste to judge, a human judges it

04 — The electric fences

The agents roam freely. The field is fenced.

Guardrails the workers cannot cross, enforced at the gate by CI, a pre-push hook, and the steward's review — and the fence line keeps growing: every incident that teaches a new failure mode becomes a new wire the next morning.

“Adherence is the worker's job; verification is the steward's.”

engineering
Reuse what exists, never duplicate it. Separation of concerns — no god files, no business logic in the UI. Type safety. No dead code.
testing
Required tests ship with the code. No zero-assertion tests. Mutation testing on the money-path services. Contract tests against the API.
design
Design tokens only, and an accessibility budget.
security
No secrets in code. Auth patterns off-limits to casual edits. Dependency vulnerabilities scanned on every PR.
deployment
CI green, the right branch, never a shortcut to prod. Sensitive code never auto-merges. Migrations validated against the live schema.
“A smarter model doesn't solve coordination — structure does. Verification solves trust, isolation solves collision, workflow solves ambiguity.”

“The agents aren't coordinated by intelligence. They're coordinated by structure.”

↗ A Self-Coordinating AI Dev Team: The Operating Model

“The parallelism was never the problem. Every meaningful failure traced to one thing — the issue tracker had quietly drifted out of sync with the code.”

↗ What broke the first time we ran three AI workers on one repo

“Adherence is the worker's job; verification is the Steward's. That second, independent check — not the agents' good intentions — is what lets the trust dial climb without the work going feral.”

↗ A Self-Coordinating AI Dev Team: The Operating Model

05 — In preview

in preview

The engineering brain, served over MCP.

Instead of every session re-reading a docs repo, the Mother CLAUDE MCP server serves the rules, the worker roster, and the incident history to a whole fleet of AI coding sessions — pointed at any docs source you own.

rules
The standards and the Definition of Done every session is held to, served live instead of copied.
roster
Stable, vendor-neutral worker identities — provenance that stays auditable across the fleet.
incidents
The failure history, so a new session inherits what already went wrong once.
Read the source ↗ MIT · Node 20+ · running in production on one fleet · npm release pending

07 — Hard-won lessons

From running it, not designing it.

01

Don't let agents share a mutable checkout — worktree isolation isn't optional.

02

A shared dependency tree is shared mutable state. Gate installs to an owner, or isolate.

03

Chat is not state. If it matters, it goes on the tracker.

04

Claim, then reconcile, then build — in that order, every time.

05

Never auto-merge work whose Definition of Done has no objective test.

06

An autonomous system needs liveness supervision that isn't one of its own sessions.