Agent Harness

demo

Test your agents

Drop a file to autofill all three, or click to upload · .json or labeled .txt/.md

1 What are we testing the agent under test · its system prompt

Drop a .txt/.md/.json file, or browse.

2 Expected results how it should behave · defines the checks

Drop a file, or browse.

3 Agent auditors the simulated testers sent at it

Drop a file, or browse.

Scenario name

Max turns / agent

Sim agents

Edited here directly, these override the parsed chips from your test description. Expertise / rules / tactics drive the same engine.

Behaviors we check (assertions)

fail = agent broke behavior

The five default assertions correspond to the guard vectors the prompt is scanned for. Deterministic assertions are JS over the agent's final state; the judge reads the transcript. Same loop as the Python engine.

Pass/fail result · latest run

Run matrix & fail rate

Run the audit a few times to see how often each behavior broke across runs.

Compare prompt versions

baseline version

compare version

Need two prompt versions in history. Run a bare prompt, add a guardrail line, run again, then compare which behaviors improved.

demo · runs locally

History

Sample data

Example runs. Your own audits replace these.

Overview

Trend

Behaviors held per run

Runs

Log

Everything that happened in this project: agents created and edited, audits run, and events your own scripts push.

Thinking…