⚖️ Governance
Agents decide and execute within recorded rules. Named people own those rules, budgets, and outcomes.
Who decides
Default
All changes can run autonomously, including public API, schema, security-sensitive, High, and Critical changes. People intervene on unresolved failures or exhausted retries. Review pauses are optional and must be requested.
The owners below are accountable, not mandatory approvers. Automated decisions record the agent identity, date, options, evidence, and consequences. They never claim human sign-off.
Decision table
| Decision | Accountable owner | Consulted on exceptions | Recorded as |
|---|---|---|---|
| What to build next inside the agreed outcome | Product Owner | Squad, customers | Scorecard row |
| Whether to run an experiment | Product Owner | Architect | Hypothesis entry |
| Implementation approach inside the squad's code | Developer | Architect | Pull request |
| Public contract or cross-service boundary change | Architect | Owning squads | Decision record |
| Data model change affecting other squads | Architect | affected squads | Decision record + migration plan |
| Security-sensitive change (auth, PII, billing) | Architect | Developer, squad | Review record on the pull request |
| Adding a new runtime, database, or vendor | Architect | DevOps | Decision record |
| AI usage policy and model access | Architect | Legal | Policy version |
| Adopting a new model, agent, or prompt standard | Architect | Squads | Evaluation report |
| Release of a customer-facing commitment | Product Owner | Architect | Changelog |
| Rollback during an incident | DevOps or QA and maintainer | None — act first | Incident record |
| Deleting an invalidated feature | Product Owner | Squad | Ledger verdict |
Escalation
Escalate only for: cross-squad conflict, security exposure, compliance obligation, external commitment, or budget impact. An escalation names the decision, the options, the recommendation, and the date by which silence means the recommendation is adopted.
Blast radius and required automated validation
Blast radius means the scope and reversibility of a change. Think sets the tier; Plan checks it. The tier determines the evidence required in Prove, not who must sign.
| Blast radius | Criteria | Required automated validation |
|---|---|---|
| Low | Internal refactor, non-breaking schema change, copy edit | Green controls and independent assessment of intent, behaviour, and safety |
| Medium | New public endpoint, schema addition, external integration | Low-tier evidence plus contract and integration checks appropriate to the change |
| High | Auth, permissions, billing, PII, data migration | Access-boundary and safety checks, with verified rollback evidence |
| Critical | Irreversible migration, breaking public contract, infrastructure change | High-tier evidence plus a decision record, continuity checks, and a validated staged rollout |
Line-by-line review ban
Manual line-by-line code review is not an allowed gate. Use automated checks and controls. Prove independently assesses intent, behaviour, and safety. Human review is optional or handles exceptions.
Accountability
AI transcribes meetings, extracts needs, writes code, and fixes failed checks. A named person remains accountable, without approving each run, for:
- the objectives and policies under which changes are accepted automatically,
- any customer-facing commitment, and the decision of what problem to solve,
- the controls and evidence used to validate business intent and safety invariants,
- security, compliance, and architectural boundaries, including automated High/Critical decisions,
- any change to models, prompts, or agents affecting trust or compliance,
- the resolution of any incident, and its postmortem actions.
Tooling
Only tools on the sanctioned list may be used with company or client data. Sanctioning requires: enterprise tenancy, no training on submitted data, audit logging, SSO, and a signed data processing agreement.
Auditability
| What | Retained | Where |
|---|---|---|
| Agent actions on repositories and infrastructure | 12 months | Platform audit log |
| Model, prompt, and agent versions used in CI | 12 months | Build metadata |
| Evaluation reports for adopted changes | Indefinitely | Repository |
| Cost per squad and per workflow | 24 months | Cost dashboard |
Cost
Model spend is attributed per squad and reviewed weekly. A workflow whose cost exceeds the value it creates is redesigned or retired — cheaper models, tighter context, or no AI at all. See Cost Management.
Violations
A policy violation is handled as an incident. A policy that can only be respected through vigilance is a Deterministic Controls defect.