> ## Documentation Index
> Fetch the complete documentation index at: https://docs.zerotwo.ai/llms.txt
> Use this file to discover all available pages before exploring further.
# Subagents
ZeroTwo can spawn specialized subagents during Work, Cowork, and ZeroCode runs. Subagents handle parallel exploration, implementation slices, or research, then return summaries to the main thread.
## Where to manage agents
Open **Customize → Agents** in ZeroTwo to browse specialized agents ZeroTwo can spawn:
* **Installed** — agents available to your account
* **Personal / Plugin / System** — filter by source
* **Explore** — discover agents from plugins you can install
Each agent has a name, description, and optional icon. Plugin-owned agents appear when you install the parent plugin.
## How subagents show up in a run
When the main agent delegates work:
1. ZeroTwo starts one or more subagent runs.
2. Activity appears in the Work / Code panel and as viewer tabs you can open.
3. Each subagent reports progress, findings, or failure back to the main thread.
4. The main agent synthesizes those results into the final response.
Ask ZeroTwo to delegate independent parts of a task when work can run in parallel — for example, exploring several packages, running test suites, or drafting separate sections of a plan.
## Settings that affect subagents
In **Settings → Advanced** (and related desktop settings):
| Setting | What it does |
| --------------------------- | -------------------------------------------------------------------------- |
| **Max Agent Steps** | Caps steps for a parent Work, Cowork, or ZeroCode run |
| **Max Sub-agent Steps** | Caps steps for each delegated subagent run |
| **Sub-agent Model Routing** | Lets the agent route hard, medium, and easy sub-tasks to configured models |
| **Auto Compression** | Compacts long Work, Cowork, and ZeroCode runs as context grows |
Desktop permission mode and sandbox settings still apply to tools subagents use. See [Permissions](/permissions) and [Sandboxing](/sandboxing).
## Why use subagents
Even with large context windows, dumping every exploration note, test log, and stack trace into the main chat can hurt quality over time.
Subagents help by:
* Keeping the **main agent** focused on requirements and final outputs
* Running specialized work **in parallel**
* Returning **summaries** instead of raw intermediate noise
They use more tokens than a single-agent run because each subagent does its own model and tool work.
## Project guidance for delegation
Use [`AGENTS.md`](/agent-configuration/agents-md) (or `ZEROTWO.md`) to tell ZeroTwo when to delegate — for example, “explore each package with a subagent before proposing an architecture change.”
Skills and plugins can also define agents ZeroTwo may spawn. See [Skills and plugins](/skills-and-plugins) and [Build plugins](/build-plugins).
## Related
* [AGENTS.md](/agent-configuration/agents-md)
* [Skills and plugins](/skills-and-plugins)
* [Long-running work](/long-running-work)
* [Permissions](/permissions)