When an AI agent pauses to ask a human for input, most teams treat it as a one-off failure. Ambient Code thinks that's the wrong frame — and has published a metrics framework built around the opposite premise.
The company's proposal treats every "interrupt" — any moment an agent halts because it's missing a decision record, policy rule, or capability — as a structural signal. Each category maps to a specific fix: missing context gets an Architectural Decision Record, a policy gap gets a new rule added to the system's governing "constitution," a tool failure gets a skill patch. Once a fix lands, that whole class of interrupt disappears. Repeat enough triage cycles and interrupt rates compound downward.
The framework adapts <a href="/news/2026-03-14-8-levels-agentic-engineering-framework">DORA metrics from DevOps</a> — the framework's author cites direct prior experience applying DORA in cloud services engineering — into five tracking signals for agentic systems:
- Interrupt Rate: interrupts per agent-task (target: down) - Autonomous Completion Rate: tasks completed with zero human input (target: up) - Mean Time to Correct: time from interrupt to human response (target: down) - Context Coverage Score: percentage of interrupt categories with a structural fix in place (target: up) - Feedback-to-Demo Cycle Time: time from feedback signal to working demo (target: down)
Pull Request #51 is where the framework stops being abstract. Ambient Code's monitoring process identified that its own ambient-code[bot] agent had missed the Error terminal state in React Query polling logic three times in a single session. Rather than flag it for manual review, the system opened a pull request patching the prompt gap. A human merged it. The company cites this as the correction-to-PR workflow already running in production — not a hypothetical.
The reference library of GitHub Actions patterns includes automatic conversion of labeled issues into draft pull requests, parallel agents running separate security, bug, and code quality reviews before any human sees the output, and a self-review reflection step where the agent critiques its own work before handoff. That last pattern is aimed squarely at the <a href="/news/2026-03-14-spec-driven-verification-claude-code-agents">redline cycle</a> — the back-and-forth that typically burns time between an agent's first pass and a human's approval.
Ambient Code has not yet published live performance data from its proposed weekly triage meeting format, but says it will once numbers are collected.