What to Measure for AI Agents in Production: Five Metrics and How to Set Targets
Measure outcomes, not uptime: task success rate checked against the end state, escalation and intervention rate, policy-violation rate by severity, cost per completed task, and time to verified resolution. Express each as good events over valid events, set targets the way SRE teams set SLOs, agree in advance what happens when a budget runs out, and treat every severe violation as an incident.
On this page
Many agent dashboards answer one question: is the agent running? Requests per minute, response times, error codes, tokens spent. These are the numbers infrastructure has always produced, and you need them.
They do not say whether the agent did its job. Google's SRE book already names the gap. Its definition of errors, one of the four golden signals of monitoring, includes the implicit kind: "an HTTP 200 success response, but coupled with the wrong content". Only end-to-end tests, it adds, can detect that you are serving the wrong content.
For an agent, most failures are 200s.
Which metrics matter for AI agents in production
The SRE workbook's chapter on implementing SLOs recommends treating an indicator as "the ratio of two numbers: the number of good events divided by the total number of events". That framing suits agents, because it forces you to define both halves.
| Metric | How to count it | What it tells you |
|---|---|---|
| Task success rate | Tasks whose end state matches the goal / tasks in scope | Whether the job got done |
| Escalation and intervention rate | Tasks handed to, or rescued by, a person / all tasks | How much of the autonomy is real |
| Policy-violation rate | Tasks with a violation, by severity / tasks evaluated | Whether the job was done within the rules |
| Cost per completed task | Total spend / successful tasks | What a result actually costs |
| Time to resolution | Tasks resolved within target / tasks resolved | What the customer experiences |
Task success rate
Count success by the state of the world, not by the agent's account of it. τ-bench, a benchmark for agents that use tools under domain policies, scores a conversation by comparing the database state at the end with the annotated goal state. Do the same in production where you can: was the refund issued, the ticket closed, the record updated? Where you cannot, sample and review.
τ-bench also proposes pass^k: the chance that the agent succeeds in all of k independent trials of the same task. In production, doing it right every time matters more than doing it right once.
Escalation and intervention rate
Count these separately. An escalation is the agent handing a case to a person because a rule says it should. It is planned, and a sign the routing works. An intervention is a person stepping in to stop, fix or undo something the agent did. It is unplanned, and a sign a failure got through.
A falling escalation rate is good news only if interventions do not rise with it. An escalation rate of zero usually means the routing rules are not firing. And if escalations are approved without being read, the number means less than it seems; approval quality has measures of its own.
Policy-violation rate
This is the share of tasks in which the agent broke a written rule: disclosed data it should not have, committed to terms it was not allowed to, acted outside its mandate. Report it per policy and per severity, never as one blended figure, because a single overall rate hides the failures that matter.
Cost per completed task
Divide total spend by successful tasks, not by calls or attempts. Include model tokens, tool and API fees, retries and the human time spent on escalations. A cheaper model that fails more often can cost more per completed task. Spend that spirals has its own controls.
Time to resolution
Measure from the request to the verified outcome, including time spent waiting for a person. Report percentiles. The SRE book's chapter on service level objectives notes that a high percentile "shows you a plausible worst-case value", while the median "emphasizes the typical case".
How to set SLOs for an AI agent
An SLO is a target for one of those indicators. The error budget is what is left over: in the workbook's words, "100% minus the SLO". The SRE book's advice on choosing targets carries over almost unchanged:
- "Don't pick a target based on current performance." Adopting today's numbers without reflection, the book warns, can lock you into a system that "requires heroic efforts to meet its targets".
- "Have as few SLOs as possible." If quoting an SLO never wins a conversation about priorities, it is probably not worth having.
- "Perfection can wait." Start with a loose target and tighten it, rather than a strict one you later have to relax.
- "Keep a safety margin." An internal target tighter than the one you promise gives you time to react.
Two adjustments are specific to agents.
Severe violations are not a budget. An error budget makes sense for task success: some tasks will fail, and the budget says how many is acceptable. It makes no sense for a severe policy violation. Set those to zero and treat each one as an incident, with the same discipline as any other agent misbehaviour.
Decide what happens when a budget runs out, before it does. Google's example error budget policy halts changes and releases, other than top-priority bug fixes and security fixes, until the service is back within its SLO. It also requires a postmortem for any single incident that consumes more than 20% of the budget over four weeks. For an agent, the equivalent might be no prompt, model or tool changes until task success recovers, with more cases routed to people in the meantime.
Suppose a returns agent. Its owners might set, over a rolling 28 days: task success for at least 90% of in-scope requests, verified against the order system; interventions on fewer than 2% of tasks; zero severe violations; and 95% of requests resolved within ten minutes. These are illustrations, not benchmarks. Write down the reasoning behind your own. The workbook asks every SLO document to record "the rationale behind the numbers, and whether they were derived from experimental or observational data".
What NIST's AI RMF MEASURE function adds
The NIST AI Risk Management Framework puts the same discipline in a governance frame. Three of its MEASURE subcategories belong in the metric set itself, not just in a policy document:
- MEASURE 1.1: risks or characteristics that "will not – or cannot – be measured are properly documented". Say which failures your metrics cannot see.
- MEASURE 1.2: the appropriateness of metrics and the effectiveness of controls "are regularly assessed and updated". Review the SLOs on a schedule.
- MEASURE 3.3: feedback processes for end users to report problems and appeal outcomes are "integrated into AI system evaluation metrics". Complaints and appeals are a metric, not only a queue.
Our reading of the AI RMF for agents covers production monitoring under MEASURE 2.4.
Where Syntrox fits
A violation rate is only as steady as the check that produces it. Nucleus, our judge, runs at temperature 0 with a fixed seed, so the same input gets the same verdict, and every verdict is stamped with the judge model and version that produced it. Syntrox records, per evaluated trace, the score, confidence and judge version outside the agent, and its pipeline records any trace no judge could score as a failure, never as a pass.