Every growth team has a conversion rate. Almost none of them have an agent task-completion rate — and in a web where the majority of traffic is automated, that's the metric quietly deciding how much of the machine economy you capture.
Define the task, then watch an agent attempt it
A task is a goal against a real URL: find a product → add to cart → reach checkout, sign up, book a demo, submit a lead. A run executes that task the way a well-behaved agent would — following links, reading accessible names, filling fields by label — and records one honest outcome:
- Did it complete? (yes / no)
- If not, which step failed?
- Why, in a machine-readable failure class — not "it broke," but
form-field-not-machine-readableorcheckout-step-requires-human-affordance.
Aggregate that across your key flows and you get a single, defensible number: the share of agent attempts that succeed. That's your task-completion rate.
Why it beats traffic as a signal
Traffic tells you agents arrived. Task completion tells you whether they could do business with you.
A dashboard that shows "agent traffic is up 30%" feels like progress and means almost nothing on its own. Up-and-to-the-right traffic with a 40% completion rate is a leaking bucket. The completion rate is the number that maps to dollars, because it measures the exact thing an agent was sent to do.
Deterministic first, autonomous second
There's a temptation to measure this only with a "real" LLM agent reasoning its way through the flow. That makes a great demo, but it's nondeterministic and expensive — a shaky foundation for a number you want to trust release over release. The durable approach runs a deterministic suite for the trustworthy, reproducible score, and layers autonomous agent runs on top as a premium, periodic signal. Same input, same score — that reproducibility is what lets the number gate a deploy instead of just decorating a slide.
From metric to habit
The point of measuring isn't the dashboard; it's the loop. Once task completion is a number, two things become possible: you can fix the specific failures dragging it down, and you can gate regressions so a future deploy can't quietly break agent checkout again. The metric is the entry point. The fix-and-gate workflow is where it starts compounding.