openhuman/docs/DELEGATION_POLICY.md

Delegation Policy

When to delegate vs. act directly

The orchestrator follows a direct-first policy. This document codifies the four-tier decision tree the orchestrator applies to every user message.

Tier 1 — Reply directly (no tools)

Apply when: small talk, simple factual Q&A, acknowledgements, clarification requests, context already in the system prompt. Cost: 0 tokens (output only). Rule: if you can answer without calling any tool, do so.

Tier 2 — Use a direct tool

Apply when: the task needs a tool but not specialised execution (time lookup, memory read/write, cron scheduling, workspace state, listing connections). Cost: 1 tool call + parse overhead (~200-400 tokens). Rule: prefer current_time, cron_*, memory_*, memory_tree, read_workspace_state, composio_list_connections, ask_user_clarification.

Tier 3 — Delegate to a reusable async sub-agent

Apply when: the task requires specialised execution (writing code, crawling docs, running shell, calling an external integration) that the orchestrator cannot do directly. Cost: full sub-agent turn (~1-5k tokens depending on archetype). Rule: spawn the narrowest archetype that can complete the task. spawn_subagent is reusable and asynchronous by default: it first looks for a compatible durable worker for the same parent thread, agent id, toolkit/model/sandbox/action-root shape, and deterministic task key. If one is running, the new instruction is steered into it; if one is idle, it resumes from saved history; otherwise a new durable worker session is created.

Tier 4 — Explicit worker lifecycle control

Apply when: the task is long (>5 turns estimated), produces a large transcript, needs manual inspection, or the user explicitly wants it tracked/closed separately. Cost: same as Tier 3 but the parent thread is not flooded. Rule: use list_subagents, steer_subagent, wait_subagent, and close_subagent with subagent_session_id for durable control. Use fresh: true only when the prior worker is materially incompatible or the user asks for a clean worker. Use blocking: true only when the parent must synchronously wait for the child result before replying. Do not chain workers (workers cannot spawn workers).

Anti-patterns to avoid

• Spawning a sub-agent to answer a question the orchestrator already has context for.
• Delegating a tool call to a sub-agent when current_tier <= 2 applies.
• Repeatedly forcing fresh sub-agents for the same logical job, which loses worker-local context and cache affinity.
• Passing the entire parent conversation as context to a sub-agent — pass only the task-relevant slice.