Name: Reviewed Interpretation (normative): Documented invariants, deterministic builds, and review gates let an agent propose changes a human can confidently approve.
L3 is where agent-assisted development becomes genuinely dependable. The
system's invariants are documented, builds are deterministic, and review gates —
branch protection, CODEOWNERS, CI checks — stand between a proposed change and
main. An agent can read those documented contracts, propose a change, and a
human can confidently approve it because the system no longer depends on any
single contributor's discipline.
Criteria at L3
L3 criteria build on L2, L1, and L0:
DOC-070— API reference auto-generated or hand-maintainedCQV-080— Code complexity gatesTST-060— Test isolation (no shared mutable state)TST-070— Flake detection and quarantineBLD-060— Hermetic CI (no network for test/build beyond cache)BLD-070— Docker image produced and publishedOBS-050— Metrics exportedSEC-080— Security headers setTIR-040— Triage convention documentedDLM-040— CI feedback time instrumentedDLM-050— DORA metrics captured
How to climb to L4
L3 → L4 is about handing the agent the wheel. At L4 the agent can take a scoped task and drive it end-to-end — read the contracts, change the code, prove it green, and ship — with a human supervising rather than steering. That means closing the last enforcement gaps: a full CI quality gate, property-based testing on critical modules, SLOs and alerting, and structured task-intake templates an agent can act on autonomously.