Signals watches real agent traffic for interaction problems, failing tools and environmental constraints. A problematic session can then become an evaluation scenario, connecting something that went wrong in use to a test of the proposed fix.
Simulations supplies a simulated user over multiple turns in an isolated environment. Tools that write data or incur costs can be mocked, while baseline and candidate runs are compared on goal completion, knowledge retention and tool use. Scenario sets already used remain locked so results can be revisited against the same material.
That approach measures a sequence of actions rather than just the final answer. It can turn a production trace into a reproducible regression test, but the simulator and LLM judge do not replace observation of real users. The reported scores and efficiency gains are DigitalOcean's, not independent validation.
The October 12 release is a public preview for harnesses compatible with Harness Runtime, including Claude Code, Codex and OpenCode. LangChain, Hermes and CrewAI are not supported in this preview. No explicit Belgium/EEA exclusion is stated; access depends on the account, plan and service region.
