Name: Agent-driven Interpretation (normative): An 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.
L4 is the top of the agent-readiness ladder for a codebase as it exists today. A full CI quality gate enforces every contract, critical modules carry property-based or fuzz tests, SLOs and alerting describe what "healthy" means, and task intake is structured enough that an agent can pick up a scoped task and run it to completion. The human role shifts from steering each step to supervising the outcome.
At L4 an agent can take a well-scoped task, read the documented contracts, make the change, prove it green against the full gate, and ship it — escalating only when a decision exceeds its delegated scope. This is the highest rung that describes a real codebase rather than an aspiration; the rungs above it describe an agent that stands up and operates a whole business.
Criteria at L4
L4 criteria build on every level below. They are the gates that let an agent drive end-to-end:
TST-080— Property-based or fuzz testing for critical modulesOBS-060— SLOs and alertingSEC-090— SBOM generated on releaseDLM-070— Self-improving gatesCQV-090— CI enforces a full quality gateTIR-050— Structured task-intake templates
How to climb to L5
L4 → L5 is where the ladder leaves today's AI behind. At L5 the agent doesn't just change code inside an existing project — it goes from a goal to a launched product, standing up the codebase, the infrastructure, and the go-to-market. The new criteria at that rung (billing integrated, infrastructure as code) describe what a codebase needs before an agent could plausibly turn an idea into a revenue-capable business.