HarnessBench — In-Browser Agent Evaluator

1. Load Model

Model

Weights are downloaded once and cached locally. No data leaves your device.

2. Define Task

Task Prompt

Max Steps

Try:

Raw Log

▸ Live Token Stream

idle

(model output will appear here in real-time…)

Agent Trace

—

Run an evaluation to see the agent's reasoning and actions here.

Sandbox State

empty

{
  "files": {},
  "vars": {},
  "logs": []
}

Registered Actions

Harness Verification

Deterministic tests that verify the harness machinery — no model needed. Covers the action registry, sandbox isolation, unknown action enforcement, judge correctness, JSON repair, and context trimming.