Documentation
¶
Overview ¶
Example: deep_agent demonstrates the batteries-included "deep agent" harness preset (WC-A-005). A single harness.NewDeepAgent(...) call assembles the whole agent harness — planning (WC-A-001), a virtual filesystem for context offloading (WC-A-002), context-isolated subagents (WC-A-003), automatic compaction with the active plan pinned (WC-A-004), and (optionally) semantic memory recall (WC-D-001) — with a sensible default prompt and tool set, and no extra wiring.
A single deterministic mock model.Provider drives every role (the deep agent and its subagent, which inherits the same model), so the example runs with NO API keys and NO network. It works a realistic long task across two turns:
turn 1: lay out a plan with update_plan
turn 2: offload notes with fs_write, delegate research with spawn_subagent,
mark the plan complete, and produce the report
Because storage is durable (sqlite), the plan and files survive across turns, and the active plan is pinned into context so compaction never drops it.
go run ./examples/deep_agent/
Source Files