Measure AI agent cost per task: accepted work, retries and verification

A worked example: where the task becomes difficult

A fictional workspace applies an edit and then runs two repair attempts. The interface reports only the initial model call, hiding most of the task’s work. Another task uses paid runtime minutes while waiting for a repository build. Define a task identity and record stages so the operator can understand what generated, verified and retried work contributed to usage.

Decisions to make before implementation

Choose units for model usage, runtime time, storage and external operations. Record provider-reported usage where available and clearly label estimates otherwise. Avoid double-counting a job observed through several event sources. Connect cost to the actual selected model and rate period without hardcoding an unverified current price. Keep user budgets and usage views scoped to the authorized account.

A practical sequence for the work

Use the sequence below as a task boundary, not as a claim that the example has been executed. Work with approved inputs and the project’s actual architecture. If a required integration or permission is unavailable, keep that stage visibly incomplete rather than generating a plausible substitute result.

• Define task and stage identities for generation, repair and runtime work.
• Collect recorded usage with units and source references.
• Calculate totals using verified rates or clearly labeled assumptions.
• Compare cost with accepted task outcomes and identify avoidable repeated work.

A detailed brief you can adapt for your agent

Replace the illustrative context with your approved facts and controlled inputs. Keep the stated boundaries when adapting the brief. The expected deliverable matters more than a particular tool name: ask for an explanation grounded in the inspected material and evidence for the requested outcome.

Design task-level usage accounting for generation, verification, repairs and runtime stages. Preserve units and provider source, deduplicate event identity and keep estimates distinct from recorded billing. Use the selected model’s verified rate only when available. Show the cost of unsuccessful work honestly and define the approved budget stop behavior.

Failure modes that an attractive preview can hide

A completed response can consume usage even when its edit is rejected or fails. Do not equate no applied changes with no cost. Conversely, a provider dashboard may aggregate unrelated tasks. Keep attribution limits visible. Budget controls should stop or confirm additional work under the approved workflow rather than silently allowing unlimited repair loops.

Technical references: Google: Web Vitals

Acceptance checks and the evidence to retain

Keep stage usage, calculation assumptions, attribution gaps and accepted-outcome comparisons. Do not claim exact billing from incomplete telemetry.

Specific answers

Common questions