Skip to main content
Beyond watching your agent run, Rippletide can test it against realistic scenarios and propose improvements. Both need a connected agent with its runtime worker online (they do not apply to MCP servers).

​
Evaluate

Evaluation runs your agent against a set of scenarios and grades each result.
  1. On the Evaluate page, generate scenarios from your agent’s declared behavior, or write your own.
  2. Start the run. Your local worker executes each scenario against the real agent.
  3. Review the graded results on the same page.

​
Improve

Improvement is an experiment loop on top of evaluation:
  1. Rippletide proposes an improvement zone, for example a weakness in the system prompt.
  2. It prepares scenarios and a deterministic check, and shows you the plan.
  3. After you confirm, your worker runs the real agent through candidate prompt and tool variations and compares the results.
A kept candidate is a recommendation, not a deployment. Your code stays the source of truth: nothing is edited or shipped until you apply the change yourself.