How to evaluate browser agents: a practical checklist
A reproducible release protocol for browser and computer-use agents: task state, visual grounding, trajectories, side effects, recovery, security, and risk-bounded rollout.
AI agent benchmarks: GAIA, WebArena, OSWorld, and SWE-bench
A practical guide to choosing an AI agent benchmark: what GAIA, WebArena, OSWorld, and SWE-bench actually test, how to read results, and how to carry an external signal into your own release gate.
AI agent trajectory evaluation: test the path, not only the result
A practical guide to evaluating AI agent trajectories: trace contracts, tool calls, permissions, retries, side effects, graders, failure taxonomy, and release gates.
How to evaluate prompt-injection defenses: a practical checklist
A reproducible protocol for testing prompt-injection defenses: threat modeling, source-to-sink fixtures, tool traces, data exfiltration, side effects, false positives, release gates, and rollback.
How to evaluate AI tool calling: a practical checklist
A reproducible protocol for evaluating function calling and tool use: tool selection, arguments, trajectory, side effects, retries, terminal state, cost, and a release gate.
How to evaluate LLM hallucinations: a practical checklist
A reproducible protocol for evaluating factuality and groundedness: claim types, verified evidence, abstention, calibration, long-form responses, risk slices, and a release gate.