Blog.
Practical notes on testing agents in simulated environments.
Reconstruct a test environment from agent traces
Rebuild a failed Slack bot run from its trace, check where the reply lands, and verify a one-line fix. Includes a runnable Ancha scene.
How to test a Slack bot
How to test a Slack bot: mock the client, fake the Slack API, use a test workspace or run a local replica. What each catches, what it misses, and the tools.
Putting Harness² on call
We ran Google Research's Harness² on an on-call agent in Ancha worlds. With Claude Opus 4.8 as the editor it found a one-paragraph fix that took a failing task from 0 of 3 to 2 of 3.
Nothing in your tests acts back
What Ancha believes about testing AI agents: what your agent does stays done, your evals already have a home, everything that happens is an event, grade the end state, never a silent 200.