Definition
What is a subagent in a coding agent?
A subagent is a separate agent that a main agent spawns to handle part of a task, usually in parallel with others. It saves wall clock time on work that splits cleanly, such as searching several subsystems at once, and it produces interleaved output that a terminal cannot render legibly.
Last updated · Markdown version
When fan out helps
- Searching or reading across many files or subsystems.
- Reviewing one change from several independent angles.
- Any work where the parts do not depend on each other's results.
When it does not
- Sequential work where each step needs the previous answer.
- Small tasks, where the coordination costs more than it saves.
- Anything where you need one consistent view of a file being edited.
The rendering problem
Seven agents writing to one stream is not a log. A client that reads the protocol can show each subagent separately, cancel a run, or skip and retry an individual agent, which is the difference between a fan out you can supervise and one you can only hope about.
Source: Extending and automating (Muse Code docs, subagents section)
Frequently asked questions
Do subagents cost more?
They use more tokens in the same wall clock time. A per thread cost view makes the tradeoff visible.
Can I stop one subagent?
In Helicon, yes. Individual agents in a workflow run can be skipped or retried, and the run can be cancelled on its own.
Is a subagent a separate session?
It belongs to the parent thread. The session is still the thing you resume.
Keep reading
- Subagents and workflowsParallel subagents and workflow runs are hard to follow in a terminal. Helicon shows them as structure: what is running, what finished, and what you can cancel or retry.
- Background workSend a running tool call to the background, stop one, or stop them all. Cancel a workflow run, or skip and retry its agents, without abandoning the thread.
- Cost at API ratesSee what every Muse Code thread would have cost at published per token rates, broken down by day, by model and by thread. A meter reading on your subscription, not a bill.
- reasoning effortReasoning effort controls how much thinking a model does before answering. What it changes, what it costs, and when to turn it up.
- a TUIA TUI is a text user interface drawn inside a terminal. Why agents ship one first, and where the format runs out.
- llms.txtllms.txt is a plain text file at a site's root that gives language models a clean summary and an index of the pages worth reading. What goes in one.
Same Muse Code. Same subscription. Better interface.
Free and MIT licensed. Windows installer, Apple Silicon, Intel and universal macOS DMGs, and a web build against a daemon you run.