Agents
Agents that delegate to agents.
Parallel subagents, isolated context, full observability.
Why bother splitting?
A smart agent plans the work, smaller ones do the parts. It usually costs less and finishes in about the same time.
The lead agent figures out the steps so each helper has one clear job.
Use a top model for the thinking. Use cheaper models for the busywork.
Each helper's work is visible. Open any of them to read what it did and why.
The helpers inherit how you've trained the main agent: its tone, what it remembers, what skills it knows about. The answer comes back sounding like the same agent, not five different ones stitched together.
What you get
Fan out N subagents at once. Each gets its own conversation, its own model, its own context.
Run an expensive orchestrator that delegates to cheap workers. Cuts cost without losing capability.
Subagents inherit the parent agent's voice, memory, and skills unless you override.
Every subagent run is a node in the trace. Open any one to see what it did and why.
Who it's for
Orchestrator agent fans out 8 subagents to skim 8 papers in parallel; aggregates findings on return.
Code-review orchestrator spawns one subagent per file; results merged into one review comment.
How it works
- 1Enable subagent skill
Tick subagent in the agent's skill list.
- 2Describe the work
The orchestrator decides how to split it across subagents.
- 3Watch them run
The trace shows every subagent's prompt, model, output, and cost.
Frequently asked
Hand off a chunky task.
Watch one agent break it down and the helpers finish in parallel.