The evaluation gap is not a metrics problem
A VentureBeat Intelligence survey from August 2025 found that 61% of enterprises deployed AI agents or LLM features that passed internal evaluations and then caused customer-facing failures. The failure rate is not explained by bad evaluations. It is explained by evaluations that measure the wrong things.
Most evaluation frameworks optimize for task completion and per-turn accuracy. Both are necessary. Neither is sufficient. A multi-step agent that resolves 94% of test cases in a controlled environment can still hallucinate tool arguments on edge-case inputs, drift scope across a conversation, or produce plausible-looking outputs that are factually incorrect in ways the completion metric does not capture. The gap between what your evals score and what your customers experience is the highest-leverage problem in agentic AI for 2026.
This article covers the specific metrics and structural practices that close that gap, written for the engineers and QA leads building evaluation frameworks for production agents.
Why task completion misleads you
Task completion measures whether the agent finished. It does not measure how the agent finished, or whether intermediate steps were sound. For a single-turn model this distinction is small. For a multi-step agent operating across tools, APIs, and sub-agents, it compounds.
Research on pass@k reliability illustrates the problem directly: agents that score 60% on pass@1 drop to 25% on pass@8. In a lab, you run one evaluation pass and call it 60% reliable. In production, your agent runs the same task across thousands of requests under varying context, and the effective reliability is closer to 25%. Completion-based accuracy in a fixed test harness does not expose this variance.
Klarna's customer service agent handles 69% of chats autonomously across 35 languages and 23 markets, with a 2-minute average resolution against a 12-minute human baseline. Getting there required evaluating per-turn quality and escalation prediction, not just final resolution rate. Without step-level metrics, an agent that resolves the task by skipping a required verification step looks identical to one that resolves it correctly.
The four metrics that predict production failure
Trajectory accuracy
Trajectory accuracy measures whether the agent took the right sequence of actions, not just whether it arrived at a correct final state. You define an expected action graph for a task, then score each execution against it. An agent that calls a refund tool before verifying account ownership may produce the correct monetary outcome in your test environment while introducing a process control failure that regulators or customers will surface later.
Step-level evaluation against expected trajectories is the structural practice that surfaces these failures. Without it, your eval pipeline is auditing the last sentence of a paragraph and ignoring everything before it.
Scope drift rate
Scope drift measures how often an agent acts outside its defined authority, either by calling tools it was not supposed to invoke or by taking actions in a conversation that exceed what the user requested. In multi-agent systems, this compounds: an orchestrator that passes an over-scoped instruction to a sub-agent can trigger a cascade where each step is individually valid but the sequence violates policy.
Multi-agent orchestration introduces specific failure modes that single-agent evals do not expose. Measuring scope drift requires defining an activity schema, the set of tool calls and action types that are permitted for a given agent role, and validating every execution against it. ServiceNow's agentic AI deployment across ITSM, HR, and finance achieved 90% self-service resolution for IT support and 20x faster HR resolution, in part because they built evaluation-before-autonomy governance that measured per-step authority, not just outcome.
Behavioral consistency under distribution shift
Your test set represents the inputs you anticipated. Production is everything else. Behavioral consistency measures how much an agent's output quality degrades when inputs differ from training or evaluation distribution: different phrasing, unexpected user context, or novel tool states.
The 37% gap between lab benchmark scores and real-world performance is a direct measure of this degradation. Ryanair's AI agent generates 120,000 answers daily at a reported 94% accuracy rate, which required daily monitoring against benchmarks, not one-time pre-launch evaluation. Behavioral consistency is not a property you measure once; it is a property you track continuously.
Calibrated confidence versus correctness
An agent that hedges appropriately on uncertain inputs is safer than one that states incorrect information with high apparent confidence. Calibration metrics measure whether the agent's expressed certainty correlates with its actual correctness. This is distinct from accuracy: an agent can be accurate on average and still be systematically overconfident on the subset of inputs where it is wrong.
Calibrated confidence is not the same as correctness, and conflating them is a common source of customer-facing failures in support and finance use cases. Macquarie Bank's Q agent achieved a 40% reduction in security alert false positives, a result that depends directly on calibration metrics, not just resolution rate.
Evaluation as governance, not a gate
The framing of evaluation as a deployment gate, pass or fail before release, produces exactly the conditions where production failures follow. It treats the agent as a static artifact and the evaluation as a certificate. Neither is true.
Auditing agents after deployment is not a fallback for failed pre-launch evals; it is a required component of the evaluation architecture. Behavior schemas validated pre-launch need to be re-validated continuously against production traces, because model updates, prompt changes, and real-world input distributions shift agent behavior in ways no static test set anticipates.
Prefactor instruments this by recording spans at each agent step, scoring quality and risk per span, and validating execution traces against activity schemas. When a production run deviates from the expected action graph, the deviation is flagged against the audit trail rather than discovered through a customer complaint.
Detecting agent drift through schema validation is the operational practice that connects pre-deployment evaluation to post-deployment reliability. Without continuous schema validation, you are relying on customers to surface the failures your evals missed.
The organizational implication is direct: evaluation teams that report only to product ship timelines will build evaluations that confirm readiness rather than surface risk. Every ML engineer building evaluation frameworks for autonomous agents needs authority to block deployment on behavioral grounds, not just on accuracy thresholds.
Where to start
Pick one agent in production, define its expected activity schema, and run your last 500 production traces against it to count scope drift events. That count is your baseline. From there, add step-level trajectory scoring and calibration tracking before you expand evaluation to other agents.
Start evaluating your agents or read the docs for schema definition and span instrumentation.