Agentic Autonomy Telemetry
Three delivery indicators — autonomy rate, human touchpoints, and verification latency — that keep the volume of agent-generated code auditable by pairing it with specification-first governance.
The Agentic Autonomy Rate (AAR) — agent-generated lines over total merged lines — measures how much of the merged codebase governed agents synthesized versus humans typed; high-autonomy regimes target above 95%. Unlike unstructured "vibe coding," AAR is meaningful only when paired with formal verification gates and human touchpoints; it is never an excuse to bypass specification-first governance.
Two companion indicators keep the ratio honest:
- Human touchpoints — manual post-synthesis edits per merge window, driven toward zero by applying corrections upstream: amend the spec and re-synthesize rather than hand-patching the output, which forks the single source of truth (the feedback discipline of agentic linearism).
- Verification latency — wall time for multi-agent audit and CI evidence to clear, targeted at minutes rather than days.
Cost per implementation completes the picture: billed tokens plus accountable Spec Architect and gatekeeper time — counting tokens alone under-budgets ownership (compare successes per million tokens for harness-level cost normalization). When faults stem from intent drift rather than isolated bugs, weigh the re-synthesis return: regenerating a whole module from an updated spec versus patch-oriented repair (see patch outcome taxonomy).
High AAR with low touchpoints and minute-scale verification signals a specification that has earned its autonomy; high AAR without those companions is ungoverned generation, and its bill arrives as cognitive debt.
Sources
- Nguyen & Nguyen (2026), ch. 11