Completion truth
ASSAY checks the observable end state, not just the agent's final message or a hopeful plan.
ASSAY AI
Know whether your browser agent actually did the work.
ASSAY runs evidence-focused evaluations for task completion, false completion, stop-condition obedience, precision navigation, dynamic pages, and prompt-injection resistance.
Evaluation packet
Independent review
Completion evidence
Verified
False completion
Flagged
Prompt injection
Tested
Evaluation coverage
Built for agents that operate in the real browser.
Behavior boundaries
Stop-condition tests catch risky follow-through before purchases, submissions, account actions, or irreversible steps.
Adversarial pages
Dynamic content and injection prompts reveal whether page text can pull the agent away from the user's task.
How it works
A compact test loop with human-readable evidence.
1
Design the assay
Define target tasks, risky stop conditions, and the exact evidence required for a pass.
2
Run browser trials
Exercise the agent against local and live-style pages with controlled grading criteria.
3
Report the verdict
Deliver clear pass, fail, and review calls that product and safety teams can act on.
Pilot evaluations