orchestration-ladderlisted

Choose the right orchestration primitive for delegated work — a single agent, parallel agents, or a dynamic workflow — using the agent → parallel → workflow ladder, plus the spawn-vs-subagent axis (interactive-and-independent vs. harness-tracked-and-harvested), plus which MODEL/tier the worker runs on. ALSO the reference for talking to other sessions through the AIOS Glass spawn-inbox command bus. Use when deciding how to delegate a task, whether to fan work out, whether you need the result back in-session, which model or tier a delegated worker should run on, when a plan starts sprouting file-sentinels and monitor loops (the tell you picked the wrong primitive), when orienting to dynamic workflows for the first time, and whenever one session must reach another — "send a message to session X", "say hi to <session>", "message/nudge/ping a running agent", "reply to the session that spawned me", "hand this off to another session", "check who is running / which sessions are live", "kill or close a worker", or spa
The-AIOS/aios · ★ 40 · AI & Automation · score 79

Install: claude install-skill The-AIOS/aios

# The Orchestration Ladder — which primitive runs delegated work Delegating work is not one decision. It's two, asked in order: 1. **How much structure does the work need?** → climb the ladder: single agent → parallel agents → dynamic workflow. 2. **Do you need the result back in *this* session?** → the spawn-vs-subagent axis, orthogonal to the ladder. Most misfires come from answering neither and reaching for whatever primitive is top-of-mind. This skill makes both explicit. ## The ladder (lowest → highest structure) | Rung | Primitive | Reach for it when | Don't when | |---|---|---|---| | 1 | **Single agent** | One coherent task with one owner. A worker with isolated context does it end-to-end. | The work is really N independent tasks (→ rung 2) or a repeatable multi-stage pipeline (→ rung 3). | | 2 | **Parallel agents** | **2+ independent tasks with no shared state and no sequential dependency.** Different subsystems, different bugs, different research angles — investigated concurrently. | Tasks share state, must run in order, or one's output feeds the next (that's a workflow, not a fan-out). | | 3 | **Dynamic workflow** | A **multi-step pipeline** where stages have order/dependencies, you want it **repeatable and harness-orchestrated**, and the result must return to the caller. Fan-out *inside* a stage is still rung 2 — a workflow is the deterministic scaffolding around the stages. | A one-off single task (rung 1) or a flat set of independent tasks (rung 2) — a workf