Sign in
Paste the API key we sent you. It starts with hk_.
Is the harness attached to your agent?
Checking...
- Your agent sends traces
- Your agent asks for its harness by name
- A harness version is being served
How to attach your agent
- Install the SDK we sent you (Python 3.12 or later):
pip install './harness_sdk-0.2.0-py3-none-any.whl[anthropic,openai]' - Point it at this server with your key:
- Add the SDK to your agent, and fetch the harness inside each conversation so improvements reach it:
Upload past conversations
Drop your agent's chat logs here and we start learning from them right away. JSONL or JSON, in OpenAI or Anthropic message format, or a ChatGPT or Claude.ai export. Give each conversation a "model" field (such as "gpt-4.1-mini") so we can test changes on that model. Uploading the same file twice is safe.
We analyse conversations by the agent they belong to. Conversations whose file already names its agent keep that name.
Then set your agent up
We read the agent's system prompt, tools and model from its uploaded conversations and register them as its starting point, so we can test changes against it. Preview shows what we found and saves nothing.
Recent conversations
| Finished | Task type | Harness | Steps | Outcome |
|---|
Nothing yet. Conversations show up here about a minute after they end.
What we found
How your conversations went
Our verdict on each finished conversation, newest first. A verdict is "unknown" when the evidence to decide it is missing.
| Decided | Task type | Verdict | Why | Correct it |
|---|
No verdicts yet. Each conversation gets one shortly after it is listed above.
Where your agent goes wrong
Recurring ways your agent fails, found across many conversations. Each one is tracked until a change fixes it. We look every night, and as soon as one failure becomes common.
No recurring failure found yet. We look for them once enough conversations have failed.
Proposed changes
Changes to your agent's harness that we tested against your past conversations. Approve one to download it, make the change in your agent, then press "I've applied it" so the next round builds on it. If your agent fetches its harness through the SDK, approved changes can roll out on their own instead.
How changes reach your agent
No changes proposed yet.