Evaluation Loops for Tool-Using Agents

Most agent evaluations are too polite. They look at the final answer, mark it pass or fail, and miss the sequence of bad decisions that got there.

That approach breaks down once an agent plans, calls tools, retries, and synthesizes. A weak tool choice can still produce a decent answer. A strong plan can still collapse on recovery. If you only grade the output, you cannot see where drift starts.

Evaluate the Loop, Not the Snapshot

We score agent runs across four checkpoints: planning quality, tool call correctness, recovery behavior, and final usefulness. That gives us a traceable reason for every failure instead of a vague "model regression" label.

score = 0.25 * plan
      + 0.30 * tool_calls
      + 0.20 * recovery
      + 0.25 * final_output

The exact weights matter less than the discipline. A complete loop score makes it obvious whether you need better prompting, better tool contracts, or better retry logic.

Replay Real Runs

Synthetic evals help, but replayed production traces catch the ugly edge cases. We store sanitized traces with inputs, tool responses, selected actions, and outcome labels. Every model or prompt change is tested against that replay set before rollout.

This is where agent teams usually find their biggest blind spot: recovery behavior under partial failure. Timeouts, malformed JSON, and stale retrieval results rarely show up in happy-path demos, but they dominate incident reviews.

Measure Recovery Quality Explicitly

Recovery should earn or lose points on its own. Did the agent retry the right call? Did it ask the user for clarification when the system was missing data? Did it avoid repeating an irreversible action?

if tool_timeout:
  allow retry_once
  disallow side_effect_repeat
  allow user_clarification

Once you isolate recovery as a scored behavior, reliability work becomes concrete. You stop arguing about "agent feel" and start fixing specific failure paths.

Use Evals as a Release Gate

Tool-using agents need the same release discipline as any other production system. Every change should clear a minimum score, a max regression threshold, and a no-new-critical-failures rule on the replay set.

The teams shipping good agent products are not guessing more cleverly. They are running tighter loops. Evaluation is not a report card after the fact. It is the control system.