Subagents

When work decomposes into independent streams, Synapse spawns isolated subagents that run in parallel — each with its own context, tools, and lifecycle. The main thread stays clean while the workers get messy.

Why subagents

• Parallelize independent tasks instead of serializing them through one context
• Keep exploratory or messy work isolated from your main thread
• Collapse multi-step pipelines into zero-context-cost turns by writing Python scripts that call tools via RPC

Delegation in practice

Delegate via natural language ("run these three audits in parallel") or explicitly through the delegation tooling. The batch runner (batch_runner.py) supports trajectory-style batch processing for research workloads.

Good candidates for delegation

• Audits and reviews across many files or repos
• Web research where several angles can be explored at once
• Repetitive transformations applied to many independent items

Scripted pipelines

For deterministic multi-step work, prefer a script that calls tools via RPC over prompting the agent through each step. The agent is best spent deciding what to run — not re-running it.