---
title: "Capabilities"
description: "The kinds of work the agent takes on inside your workspace."
lang: en
---

Inside the workspace you choose, the agent handles the following kinds of work. Every step is visible and can be approved or rejected.

## Read and search

Search by file name or content and read files. The point is to understand a project and locate relevant context without blindly walking every directory.

## Edit files

Create or update files. Every change is recorded and leaves a checkpoint you can return to. See [Checkpoints and rollback](/docs/advanced/checkpoints).

## Run commands

Execute commands and capture the output for builds, tests and data processing. Bi asks before running them; see [Operations that need approval](/docs/configuration/permissions).

If a required program is missing, Bi reports the missing command and suggests how to install it on your platform. Installs that need administrator rights stay manual; system-level package operations (brew / apt / winget and similar) always require confirmation.

## Understand the project

List directory structure and locate function, type and variable definitions, so both you and the agent can get oriented without searching blindly.

## Parallel work

When a task splits into independent pieces, the agent spawns multiple instances to work at the same time and merges the results, cutting the wait. See [Parallel subtasks](/docs/using-bi/subagents).

## Boundaries

Whatever the operation, file access stays inside the workspace; command execution is constrained by approvals and intercept rules. More capability does not mean more permission.