agent-dispatch

name

agent-dispatch

description

Owns work placement and delegation craft across Claude Code's execution surfaces — main turn, `/btw`, fork, subagent, background task, nested subagent, agent team, dynamic workflow — with the selection economics that pick one, the delegation contract that makes a worker prompt self-contained (objective, territory, output contract, receipt), the communication topologies (star, pipeline, panel, tournament, generate-filter, loop), depth and fan-out budgeting, and the runtime mechanics of background permission surfacing, `SendMessage` resume, and worker memory. Use when parallelizing work, delegating a task to a subagent or teammate, choosing between a subagent, fork, team, or workflow, writing a `.claude/agents/` definition, designing a fan-out or review pipeline, or when a delegation stalls, over-prompts for permission, or returns weak results. Authoring runnable workflow scripts belongs to workflow-creator; gpt-5.6 offload belongs to the codex skill and Gemini offload to agy; harness configuration belongs to harness-steering.

[AGENT_DISPATCH]

Dispatch is three decisions taken in order: placement — which execution surface holds the work; contract — what the worker receives and what it returns; topology — how many workers run and how their results flow back.

Official platform facts riding any dispatch — prompting surfaces, model behavior, memory, limits, official values — resolve live, never from recall: anything Claude routes through the claudeCodeDocs MCP (mcp__claudeCodeDocs__search_claude_code_docs, mcp__claudeCodeDocs__query_docs_filesystem_claude_code_docs), anything OpenAI or Codex through the openaiDeveloperDocs MCP (mcp__openaiDeveloperDocs__search_openai_docs, mcp__openaiDeveloperDocs__fetch_openai_doc, mcp__openaiDeveloperDocs__list_openai_docs, mcp__openaiDeveloperDocs__list_api_endpoints, mcp__openaiDeveloperDocs__get_openapi_spec).

[01]-[ROUTING]

• [01]-SURFACES: per-surface mechanics — permission surfacing, resume guards, transcript persistence.