Back to Speakers

November 23–27, 2026

Nevasini Sasikumar

Independent Machine Learning Researcher

Independent researcher

Session Details

How to Evaluate an Agent That Uses Tools

An agent can finish a task and still leave you guessing about what it did along the way. A single success score won’t tell you whether it found the right evidence, chose the right tool, or recovered when the tool failed. I’ll walk through a small example that starts with a question, gives the agent a few mock documents and introduces a temporary tool error. We’ll score the answer and the evidence separately, compare two runs on the same inputs, and use the trace to tell a retrieval miss from a failed tool call. I’ll connect the exercise to what we learned building ALM-Bench: evaluation coverage changes the conclusions you can draw from a score. You’ll leave with a compact rubric and a repeatable way to decide what to investigate when an agent’s performance changes.

Full conference agenda

See every session across the AI Agents Conference November 2026 line-up.

View Agenda