Governance / intent / evidence / learning
The smallest useful
agent framework.
Give capable agents clear authority, one current goal, bounded memory, accountable roles, and independent review—without making ordinary work feel like a board meeting.
Solo by default. A workforce only when the consequence earns it.
The control plane
Five surfaces.
No hidden hierarchy.
Stable governance stays small. Volatile state is replaced. Specialist roles and procedures load only when they have work to own.
Authority, precedence, preservation, and the default solo operating model.
Always applicableOne current goal, definition of success, evidence, constraints, signals, and next decision.
Replace in placeCurated shared decisions and reusable lessons with provenance, caps, expiry, and deletion.
Bounded recallAccountability contracts and role-scoped memories loaded only for the people doing real work.
On demandRight-sized coordination for explicit, cross-system, release-critical, high-consequence, or independent-acceptance work.
Progressive disclosureA team only when needed
Accountability without theater.
- 01Frame
Freeze the outcome, evidence, authority, and observable gates.
- 02Assign
Activate only personas with an inspectable output or gate.
- 03Build
Keep write ownership exclusive and changes small enough to inspect.
- 04Verify
Exercise the real behavior on one exact artifact identity.
- 05Critique
Use a fresh reviewer who never receives the builder’s rationale.
- 06Deliver
Stop at the highest state that is both observed and authorized.
The important boundary
Build here.
Deliver deliberately.
“Fix this.”
Scoped local edits and relevant verification. No automatic push, release, or account mutation.
“Test this on my phone.”
A separate one-shot lane, pinned to an exact candidate and the app repository’s real release contract.
What stays out
Less context.
Fewer contradictions.
Swift, Unreal, web, Git, and release commands belong to the repository that can keep them true.
No raw chats, secrets, hidden reasoning, personal finance, or infrastructure identities become memory.
Recurring automations may observe and report, but they do not modify code or release builds.
Shared and persona memory have hard caps, review dates, expiry, and deletion triggers.
Portable by design
Copy the framework.
Keep the project specific.
Adopt the governing files and workforce skill, then let each repository define its own commands, architecture, verification, and delivery route.
AGENTS.md
MEMORY.md
PERCEPTION.md
PERSONA-*.md
memory/personas/
.agents/skills/the-perfect-agent/