FDLC Manifesto / 03
Autonomy is useful only when it can be trusted.
Ten principles for designing AI software factories that are bounded, inspectable, recoverable, and accountable to real outcomes.
- Operating principles
- 10
- State boundaries
- 05
- Unit of value
- Verified outcomes
These principles are the operating invariants of the Factory Development Lifecycle.
They apply whether you use Mission Control, another control plane, or your own implementation.
- 01
Explicit intent
Intent must be explicit before execution begins.
A factory cannot govern an implied objective. The outcome, constraints, owner, and proof of success must be visible before capability is delegated.
- 02
Governed plans
The Plan is a governed artifact.
A Plan is not disposable chain-of-thought or a chat summary. It is a versioned, reviewable commitment that defines scope, sequence, risk, and evidence expectations. Approval binds one exact revision.
- 03
Bounded authority
Every agent has an identity, sponsor, and bounded authority.
Capability does not confer permission. Every action must be attributable to an identity acting for a named sponsor within explicit limits.
- 04
Safe execution
Execution must be isolated, observable, and recoverable.
Autonomous work will fail. Safe operation depends on containing effects, preserving state, seeing what happened, and resuming without rewriting history.
- 05
Independent verification
Producers should not be their only verifier.
The system that created a material change shares assumptions with that change. Independent verification reduces correlated error and makes disagreement inspectable.
- 06
Evidence gates
State transitions require evidence.
An agent reporting completion is an assertion. Progression requires current, attributable proof bound to the exact subject and acceptance criterion.
- 07
Earned autonomy
Autonomy must be earned through measured performance.
Autonomy is not a product setting to maximize. It should expand only for bounded work classes with reliable verification and a demonstrated operating history.
- 08
Human judgment
Human attention belongs at consequential boundaries.
Humans are not fallback processors for routine activity. Their judgment is most valuable where risk, ambiguity, authority, or irreversible consequences are present.
- 09
Governed learning
Learning must come from validated outcomes, not merely model confidence.
Confidence and memory do not establish improvement. Learning requires attributable outcomes, reviewed signals, controlled comparison, and governed promotion.
- 10
Verified outcomes
The value of a factory is measured in verified outcomes, not generated tokens.
Generation is activity. The factory creates value only when an accepted outcome is delivered with known quality, cost, effort, and risk.
Build the system around the agent.