AgentClash

Glossary

What is agent evaluation?

Agent evaluation measures whether an AI agent completes a real task correctly under constraint. Unlike prompt tests, it scores the whole trajectory: tools, artifacts, cost, latency, and evidence quality.

live eval
gate: pass

Candidate

92correct patch, low cost

Baseline

88stable reference run

Control

73missed edge case

replay timeline

1loaded task inputs and tool policy
2ran sandbox actions and captured artifacts
3scored trajectory and validator evidence
4attached scorecard and release verdict

ci verdict

Candidate clears release gate

Correctness improved, latency within budget, and required artifacts were preserved for review.

agentclash run create --follow

How agent evaluation differs

Built for reviewable agent decisions

Prompt eval checks text from one call. Agent evaluation reruns multi-step work in a sandbox and preserves replay when something fails.

OpenTelemetry-compatible trace import

Pinned datasets and golden test cases

Baseline versus candidate regression checks

Replay trails for tool calls, outputs, and artifacts

Scorecards for correctness, cost, latency, and evidence

CI gates for prompt, model, RAG, and tool changes

Workflow

Typical eval workflow

Import the evidence

Start from OpenTelemetry traces, curated datasets, support transcripts, or a real failure your team already saw.

Pin the baseline

Record the current accepted behavior so every prompt, model, RAG, or tool change has a fair comparison point.

Replay the evidence

Inspect tool calls, outputs, artifacts, latency, cost, and judge evidence when a candidate gets worse.

Gate the release

Compare candidate and baseline runs, then fail CI before a regression reaches users.

FAQ

Agent evaluation FAQ

Is agent evaluation the same as LLM benchmarking?

Benchmarks compare models on fixed tasks. Agent evaluation also covers your prompts, tools, harness, and release gates on workloads you own.

What outputs does an agent evaluation produce?

A scorecard, replay of the trajectory, artifacts from the run, and a pass or fail against validators and gates you define.

Where should teams start?

Promote one escaped failure into a challenge pack, establish a baseline run, then compare the next candidate in CI or a benchmark eval.