Isolated Subagent Fan-Out
Each parallel unit runs in its own subagent context, so wide fan-outs never degrade a single long context. One agent per unit means independent work scales horizontally instead of serially.
Skill
Divide and Conquer
Break big jobs into parallel, subagent-driven task graphs
Use in Claude Code
/amplify:divide-and-conquer Divide and conquer turns a large job into a folded task graph of typed nodes — agent, fn, expand, and switch — executed on the amplify engine. Independent work fans out across isolated subagent contexts while values pass by reference so payload bytes never flood the orchestrator. It supports both human-reviewed decomposition of bare prompts and direct execution of skills that ship their own authored graphs.
Key Insights
Typed Folded Task Graph
Graphs compose from four node kinds — agent for non-deterministic work, fn for pure transforms, expand for fan-out, and switch for branching. Every node carries an explicit output_schema and the engine validates the whole shape before running.
Values by Reference, Never Bytes
The orchestrator only handles ids and handles; payload bytes stay out of its context. Cross-node data flows through the value store, keeping the orchestrator lean and the contexts clean.
Two Entry Modes
Mode A decomposes a bare prompt into a draft graph for human review and editing before running. Mode B loads and runs a skill that ships its own authored graph plus helper modules, with no review gate.
How It Works
1
Decompose or Load
Spawn a sub-agent to draft a graph, or load a skill's shipped graph.json.
Mode A spawns one decomposition sub-agent that turns the bare prompt into a draft folded-graph declaration with a real expand fan-out and gather reducer. Mode B loads an authored graph plus its fn modules directly.
2
Validate the Graph
Validate against the task-graph schema and dry-run init.
Write the draft to a temp file, validate against task-graph.schema.json, and dry-run task.mjs init to exercise the engine's validateGraph — rejecting unknown types, dangling deps, and missing output_schema.
3
Human Review and Edit
Present the validated draft for plan-mode review before running.
Surface the fan-out width, per-unit work, and gather step via the plan gate. The human may edit nodes, prompts, the unit list, or schemas before approving. Re-validate after any edit. Mode B skips this gate.
4
Run the Scheduling Loop
Drive the graph through the execute-plan scheduling loop.
init captures the GRAPH_ID, then ready / dispatch agent in the background / exec-node for fn, expand, and switch, honoring the concurrency window and the single-writer commit lock.
5
Gather Results
Reduce per-unit outputs with an all-resolved gather fn.
An fn reducer with require all-resolved gathers successes even when some children fail, so partial results always collect and one failure never sinks the run. Present the gathered result to the user.