Swarm

When work can be divided, swarm lets multiple agents handle different parts and return their findings to one conversation. You describe the goal, inputs, and expected deliverable without manually opening a separate conversation for every workstream.

For a release-guide review, one agent can check feature descriptions while another checks instructions. Their findings still need to be combined, deduplicated, and reconciled. Useful division of work matters more than the number of agents.

Choose a starting point

Start with a read-only review. Its inputs and findings are easy to compare, and you can judge whether the division of work makes sense.

Distinguish controls from execution

You do not need to enable everything to start. Complete one ordinary read-only swarm task, then add work or connect devices as needed. The corresponding guides hold submission limits, recovery procedures, and device-state details.

Decide what can run in parallel

Independent work has clear boundaries: analyze separate sources, inspect different modules, or review one document from different perspectives. Each workstream should be able to progress independently, with a clear way to combine its output.

Dependent work needs an order. Agree on data definitions before calculating metrics; finish chapters before checking their references. Run the independent parts together and leave dependent work until its inputs are ready.

A short rewrite, simple question, or operation requiring your approval at every step belongs in an ordinary conversation. Parallel work adds model calls, context overhead, and integration effort.

Give swarm a clear goal

Describe what each workstream should read, do, and deliver:

Use swarm to review this release guide. Have one workstream check feature descriptions and another check instructions. For each issue, cite its location, explain the problem, and suggest a change. Combine duplicates and retain disputed findings. Do not edit files yet.

Deduplication and retaining disagreement are part of the deliverable. Several complete-looking answers do not yet constitute an integrated result.

Separate conversations from devices

The conversation card shows work on the current task. The device panel shows connections, experience exchange, and delegated activity. Local parallel work does not require pairing another computer and does not create a remote device entry.

Experience exchange and permission to execute tasks are authorized separately. Pairing does not enable everything.

After the work finishes

First check whether the deliverable solves the original problem. Then decide which methods are worth retaining. Swarm organizes this run; Reusable experience helps reuse validated methods later. Finishing a task does not by itself validate a reusable method.

These guides describe desktop workflows. Ordinary swarm, continuous swarm, and device collaboration have separate controls; some capabilities depend on the connected runtime and organization settings.