2 minute readNewAgent-ready
Test how your email agent behaves before it talks to anyone
Behaviour tests send your email agent a set of test emails, from offers and refusals to pressure and attempts to get private details out of it, and show how it handled each one. Nothing is sent or saved, and runs use your own model key.
An email agent that negotiates or pays on your behalf should be tested before it meets a real counterparty. Behaviour tests do that from the agent's page.
- A pack for each kind of agent. A negotiation agent gets offers below and above its target, a decline, pressure to accept on the spot, a question about the budget, a request for the buyer's private details, a request to send its summary to a new address, and one email asking for two things at once. Payment and general agents get packs of their own.
- The real agent, nothing real sent. Each test email goes through the agent's own instructions, model and tools. Every email it tries to send and every offer it tries to record is captured and checked instead. Prices come from the agent's own deal when it has one.
- Plain results. Each scenario passes or fails with the checks behind it and what the agent tried to do. A scenario the model could not be reached for says so and never counts as a pass. A new run says which scenarios got worse or were fixed since the last one.
- Watch it run. Every scenario is a row that shows as queued, running or finished, with a clock for the run and the time each one took. Finished scenarios open straight away. Cancel a run part-way, or re-run only the scenarios that failed.
- Try a draft prompt before saving it, and name words the agent must never reveal.
- Agents test themselves with their own credential, and the workspace owner's key can run tests too.
Runs use your workspace's own Anthropic key, up to 20 a day, and each run shows what it cost.