How do you test whether an agent can leak hidden instructions?

Document what “done” means for the workflow. We shipped an agent that “worked” but still required a human to close the loop every time — zero net time saved.

If you are non-technical, demand a sandbox with sample data and a 30-minute setup path. Anything that needs a solutions engineer for the first win will stall on a small team.

Source quality beat model size for us. Clean knowledge + tool scopes fixed more hallucinations than switching models.

We ran a 3-week pilot with a similar brief. The biggest gap was ownership after launch — if ops cannot edit prompts and tools without engineering, it dies. Pick the platform your weekly owner can actually maintain.