Guardrails¶
The model only works if guardrails prevent agent drift from contracts. Without enforcement, the boundary exists in documentation but not in practice.
Core Guardrails¶
| Guardrail | Mechanism |
|---|---|
| Event schema conformance | Generate typed classes from schema definitions; type errors = contract violations |
| SLA↔SLO traceability | Validation script checks every SLA has a matching SLO definition and a burn rate alert |
| ADR coverage | Validation script checks no service pattern is deployed that contradicts an accepted ADR |
| Metric naming | Observability requirements doc mandates metric names; SLO queries depend on them |
| Contract ownership | CODEOWNERS maps contracts/ to senior engineers; no agent-initiated PR may modify contracts |
The Validation Script¶
The validation script is the primary enforcement mechanism. It runs on every CI push and:
- Fails if implementation exists without a corresponding contract
- Fails if a SLA file exists without a matching SLO definition
- Fails if a SLO file exists without a matching burn rate alert
- Fails if a service is deployed with a pattern that contradicts an accepted ADR
This script is itself produced by the Script Authoring Agent (Persona 3) from a task spec. It runs in CI via the Script Executor (Persona 4).
Contract Immutability¶
No agent-produced PR may modify files under contracts/. CODEOWNERS maps the contracts directory to senior engineers who must approve any change. This is the hardest guardrail — it makes the boundary structural, not cultural.
PR Review as Guardrail¶
Agents produce PRs; humans review them. The review question is always: does this artifact correctly express the contract? Not: is this idiomatic code?
This reframe is important. The reviewer is not judging the agent's style — they are verifying the agent's fidelity to the contract. A reviewer who cannot answer "which contract authorised this line of code?" should request changes.