Verified facts
- Signals classifies interaction, tool-execution and environmental problems, then lets a session become an evaluation scenario.
- Simulations runs a simulated user across multiple turns in an isolated environment; tools that write data or incur costs can be mocked.
- Baseline/candidate comparisons cover goal completion, knowledge retention and tool use; used scenario sets remain locked so results can be revisited.
- The preview supports harnesses compatible with Harness Runtime, including Claude Code, Codex and OpenCode, but not LangChain, Hermes or CrewAI.
What changes
Agent quality also depends on the sequence of actions, not just the last answer.
Practical implication
Look for recurring failures rather than a demo score.
Availability
Belgium/EEA: the source specifies no explicit regional exclusion; availability depends on the account, plan and service region.
Limits and uncertainty
Scores and efficiency gains are DigitalOcean’s; an LLM judge and simulator do not replace observation of real users.