# Athas Agent

Source: https://athas.dev/docs/athas-agent

Use the built-in coding agent with project context, tools, permissions, and reviewable edits.

Athas Agent is the built-in coding agent. It can answer questions, inspect your workspace and
use supported tools to make changes. Its model and connection come from your AI settings.

## Start a task

1. Open the AI panel or choose **New Athas Agent** in the command palette.
2. Select a model with the capabilities your task needs.
3. Describe the expected behavior, relevant files and constraints.
4. Attach a selection with **Add Selection to Agent Chat** when it helps explain the task.
5. Inspect tool requests and review the resulting changes.

For example: “Find why this test fails, fix the cause, and run the relevant suite. Preserve
unrelated edits and summarize any remaining failure.”

## Models and context

Choose the default model in **Settings > AI**, or override **New chats** under **Other features**.
See [Providers and Models](https://athas.dev/docs/ai-providers) for keys, Ollama and custom endpoints.
[Agent context and instructions](https://athas.dev/docs/agent-context) explains project rules and excluded paths.

## Permissions and review

Review the operation and scope when Athas asks for permission. Read a file diff before accepting
an agent edit and inspect command results before assuming tests passed. A provider's capabilities
and your organization policy can limit the tools or models available.

**Follow Agent** controls whether the editor follows files the agent opens. **Steps before pausing**
limits model requests per turn before asking to continue. Both are in AI agent settings.

## Extend or share a session

[MCP servers](https://athas.dev/docs/mcp) can add tools. [Share to Web](https://athas.dev/docs/sharing) creates a read-only snapshot;
private cloud session sync is a separate opt-in feature. [External agents](https://athas.dev/docs/external-agents)
run their own coding CLIs, while [Cloud Agents](https://athas.dev/docs/cloud-agents) run repository tasks in a hosted sandbox.