Subagents and parallel work

A subagent is a separate, clean-context AI run that does one focused job (research, a lookup, a multi-step task) and returns a short summary. It keeps the main chat tidy and lets CX work on several things at once.

How it is used#

• Ask for parallel work ("compare these 5 laptops, research each in parallel") or let CX decide.
• subagent_dispatch runs one task; subagent_dispatch_parallel runs 2 to 6 independent tasks at the same time and returns all summaries together.
• A subagent does not see your conversation, so CX restates the context it needs and gives each run a short label you can recognise.
• Long runs can go to the background; CX checks them with subagent_get and lists them with subagent_list. subagent_cancel stops one.

Limits#

• Each assistant can run 3 subagents at once by default (configurable 1 to 8).
• There is a global cap of 30 running subagents across all assistants.
• Over the cap, the dispatch fails with a clear error and CX waits for a slot.

Approval#

Every dispatch needs your confirmation (it spends model quota). You can choose Always allow if you trust CX to delegate freely.

Named profiles#

If named subagent profiles are configured, CX can pass a profile name to use a specialised helper.