Subagents

Fan work out to focused agents with their own context and budget caps, then collect just the answer.

What a subagent is

A subagent is an isolated agent loop spawned by the task tool. It gets its own context window, its own tool set, and its own budget, and returns a summary to the parent — not its full transcript.

That isolation is the point. Searching a large repository can burn a lot of context on material the main session will never need again; doing it in a subagent means the parent receives the conclusion and none of the file dumps. Up to 20 subagents can run in parallel.

Built-in presets

Five presets cover the common shapes, each with a model tier and tool set matched to the job:

Preset	Tools	Purpose
`planner`	Read-only	Read the codebase and return a concrete implementation plan.
`reviewer`	Read-only	Review changed files and return severity-tagged findings.
`build-fixer`	Read, write, bash	Run the build, diagnose failures, apply minimal fixes.
`explorer`	Read-only	Locate code and cite file paths and line numbers.
`custom`	All	Inherits the parent model; follows whatever prompt you give it.

Note that three of the five are read-only. A subagent that cannot write is a subagent whose worst failure mode is a wasted call — which is why exploration and review are the highest value things to delegate.

Budgets and limits

Every subagent runs under caps on steps, tokens, cost, wall-clock timeout, and the size of what it returns. A subagent that goes in circles hits a ceiling and reports back instead of running up a bill.

Custom agents

Define your own presets for work your team repeats:

bash

tachyon agent list
tachyon agent create

A custom agent pins a system prompt, tool set, and model, so “run our migration checklist” becomes one delegated call rather than a paragraph of instructions retyped each time.