Snapshot
A correct answer is not enough to establish that an agent behaved correctly
An AI agent does more than generate an answer. It finds information, selects tools, requests actions in external systems, and sometimes returns a decision to a person. A correct final sentence can therefore conceal a wrong customer lookup, an unnecessary write attempt, or a skipped approval. Public material available through July 11, 2026 shows OpenAI, Anthropic, Google Cloud, and AWS all building evaluation and observability surfaces that include tool use and execution paths alongside answer quality.
28
Primary sources
Official documentation and papers avoid relying on a single vendor’s framing.
3
Quality contracts
Measure the final answer, execution path, and operational safety separately.
Trace
Shared evidence
Tool calls and intermediate results connect evaluation with incident investigation.
Not one score
Decision surface
The goal is to identify which contract regressed, not to optimize a composite number.
What Changed
The object of evaluation is expanding from the answer to the way work is carried out
OpenAI’s agent evals and Agents SDK tracing treat generations, tool calls, handoffs, and guardrails as a record of execution. Google Cloud publishes agent-oriented metrics that include final-response quality, tool-use quality, hallucination, and safety. AWS AgentCore Evaluations works over sessions, traces, and spans, and can accept an expected response, goal assertions, and an expected tool trajectory as evidence. The unit of evaluation is therefore moving beyond a single prompt-response pair toward the progress of a stateful piece of work.
Observation
Measuring an execution path does not mean grading a model’s private chain of thought. It means comparing observable tool choice, arguments, ordering, approvals, outcomes, and recovery behavior with the expectations of the business workflow.
Three Contracts
Separating three quality contracts makes evaluation more useful for operational decisions
1. Final-answer contract
Measure whether the user-facing result is accurate and meets the required format, evidence, and tone. Reference comparisons, deterministic checks, human review, and rubric-based model judging can all contribute. The central question is what the agent returned.
2. Execution-path contract
Measure whether the agent selected appropriate tools, used them in the right order, avoided unnecessary actions, and stopped or recovered correctly after failure. In a booking change, identity verification, availability lookup, proposal, approval, and write should be visible as a sequence.
3. Operational-safety contract
Measure whether latency, error rate, cost, guardrail trips, approval waits, and sensitive-data capture stay within an acceptable range. Passing an offline answer and path test is not a shipping guarantee if production shows loops or attempts beyond the agent’s authority.
Evidence Loop
Dataset evaluation and production observation should use one connected stream of evidence
A development evaluation set should include more than normal requests: ambiguous instructions, missing inputs, insufficient permissions, tool outages, and actions that require approval. For each case, record not only an acceptable final result but also required or prohibited tool actions and expected stopping conditions. OpenAI traces, Google Cloud evaluation results with traces, and AWS spans with evaluation results point in the same direction: do not detach the score from the execution record that explains it.
Check contracts with fixed cases
Use core workflows, edge cases, and forbidden actions to catch answer and trajectory regressions.
Inspect paths in limited live use
Limit the agent to approved users or low-risk operations and inspect traces, failures, and approval waits.
Monitor evaluation signals as operational signals
Track tool failures, retries, latency, and stop rates as well as scores, then feed reviewed failures back into the evaluation set.
Scenarios
Execution-path evaluation matters where work crosses an external action or judgment boundary
Customer operations
Check not just the wording of a refund explanation, but whether the agent looked up the right order, verified eligibility, and requested approval before executing a refund. Fluent prose does not compensate for an incorrect action.
Internal research
Alongside summary accuracy, check that only authorized documents were used, unsupported claims were avoided, and a search failure was not filled with a guess.
Development assistance
Even a change that passes tests should be evaluated for unrelated file edits, unapproved dangerous commands, and whether the agent can stop with an inspectable diff when a task fails.
Design Implication
Evaluability becomes a design condition for the agent, not a dashboard added later
Start by naming business-critical actions and making tool calls and approval results traceable with correlation IDs. Then define separate shipping thresholds for answer quality, path quality, and safety rather than forcing them into one score. A helpful answer is not a reason to permit an unapproved write. Finally, review and sanitize real production failures before returning them to the evaluation set. This loop makes it easier to detect quality loss when the model, instructions, tools, or policy changes.
Takeaway
Agent evaluation is shifting from selecting plausible answers to confirming that externally consequential work can proceed, stop, and be explained as expected. Evaluation and observation of the execution path provide the foundation for iterating on new capabilities safely.