v3 · architecture
Hierarchical
a lead spawns sub-agents on the fly
Hierarchical architecture
v3 hierarchicalA lead agent decomposes the task at runtime, spawning sub-agents for each part — and those sub-agents can spawn their own, forming a tree (depth-capped). Children run in parallel; each parent synthesizes its children's results up to a final answer. Best for open-ended tasks that break into nested subtasks.
Lead
Research lead
sub
sub
Build lead
sub
sub
lead spawns sub-agents at runtime · children run in parallel · depth-capped
How it works
- 1A lead agent decomposes the task into sub-tasks at runtime.
- 2It spawns a sub-agent per part; sub-agents can spawn their own children (depth-capped).
- 3Sibling children run in parallel; research-y leaves get a real web-search tool.
- 4Each parent synthesizes its children’s deliverables up to a final answer.
When to use
Open-ended tasks that naturally break into nested, independent subtasks.
Trade-off
Emergent tree shape is less predictable; recursive spawning + synthesis costs more tokens.
Agents
LeadDecomposes the task and synthesizes results
Sub-agentSpawned dynamically with an invented role
Note
Recursive decomposition shines on open-ended work, but the tree shape is emergent — cap depth/width and watch cost, since every parent adds a synthesis pass.