> ## Documentation Index
> Fetch the complete documentation index at: https://docs.strawberrybrowser.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Agent runs

> Understand agent work, work traces, sub-agents, feedback requests, interrupted work, and how to review what Strawberry did.

An agent run is a companion doing work, not just replying. It may search, open tabs, read web pages, use connected apps, create files, run code, ask for feedback, or spawn sub-agents.

## What agent runs can do

Depending on permissions and available tools, a companion can:

* Search the web
* Read and compare pages
* Open and navigate tabs
* Use connected app tools
* Read and write workspace files
* Create reports, spreadsheets, documents, and artifacts
* Ask follow-up questions
* Run sub-agents for scoped investigations

## Work traces

Strawberry shows work traces so you can inspect what happened. A work trace can include:

* Searches
* Tool calls
* File operations
* Browser actions
* Sub-agent activity
* Questions and feedback requests
* Progress from a `PLAN.md`

Use the work trace to audit sources, understand decisions, and resume from partial work.

Completed connected-app actions may include a short **Go to ...** button that opens the created or updated item, such as a document, spreadsheet, email, draft, or calendar event.

## Sub-agents

Sub-agents help split complex work into smaller tasks. For example, a market research task might spawn separate sub-agents for competitors, pricing, customer reviews, and source validation.

Each `strawberry.startAgent()` sub-agent can get its own `PLAN.md` in that sub-agent's session folder.

## Feedback requests

Companions can ask for feedback when they need a decision, permission, missing data, or human judgment. Some feedback requests appear in the chat, sidebar, or right-side panel.

## Interrupted sub-agents

If a sub-agent is interrupted, Strawberry can show a banner so you can reopen or resume the thread. Review the context before continuing, especially if the task involved browser state, credentials, or an external app.

## Best practices

* Give clear success criteria before starting a broad agent run.
* Ask for source links and confidence levels when research quality matters.
* Review work traces before relying on a generated file.
* For sensitive workflows, say what the companion may do automatically and what requires approval.
* For long tasks, ask Strawberry to create a plan first.