DeepEval Agent Trajectory CI Testing: Task Completion + Tool Correctness + Step Efficiency as Regression Gates

DeepEval by Confident AI·high signal

DeepEval v3.0 ships agent-specific metrics that evaluate full multi-step trajectories—not just final answers—including task completion rate, tool correctness (right tool, right parameters), and step efficiency (did the agent reach the goal without unnecessary detours). Wiring these metrics into CI as regression gates means a refactored agent that passes all output tests but increases average tool calls by 40% will fail the pipeline before it ships. This is the practical implementation of the 'trajectory metrics for process debugging' principle without building your own harness.