deepagents-harnesslisted

# Deep Agents Harness Design long-running agents as a harness with explicit state, bounded delegation, durable artifacts, and reviewable tool calls. Use a higher-level harness when the work needs more than one ReAct loop. ## Use When - The agent must plan, delegate, write files, run tools, and resume after interruption. - Tool output is too large for the context window and needs filesystem offload. - Subagents need separate context windows or specialist instructions. - Humans must approve, edit, or reject high-risk actions. - You need production traces, checkpoints, evals, or deployment hooks. ## Harness Design 1. Define the persistent state object: task, plan, artifacts, decisions, approvals, and final output. 2. Split capabilities into tools, subagents, filesystem, memory, shell, and human review. 3. Give each subagent a narrow contract: inputs, allowed tools, expected artifacts, and stop condition. 4. Store bulky observations as files and pass summaries plus paths through the state. 5. Add approval gates for network calls, shell commands, writes outside workspace, credential access, deletes, and external side effects. 6. Trace every step with state diffs, tool inputs, tool outputs, and handoff reasons. 7. Evaluate the trajectory, not just the final answer. ## Decision Table | Need | Pattern | | --- | --- | | Short single-step assistant | Plain tool-calling agent | | Stateful workflow with branches | LangGraph-style graph | | Long-horizon work with files and subagent