How we built an Evaluation Workflow for a Deep Agent with Harbor and LangSmith

Deep Agents provide a framework for quickly building AI agents with built-in planning, filesystem tools, and subagent capabilities. This guide demonstrates a systematic evaluation workflow using Harbor for sandboxed execution, Terminal Bench 2.0 for benchmarking across 89 real-world tasks, and LangSmith for observability and trace analysis. The workflow establishes baseline performance (42.65% on Terminal Bench), identifies optimization opportunities through trace analysis, and enables data-driven improvements like reducing environment setup latency by pre-populating context information in prompts.