What the AI agent does
Autonoly's agent is an AI that works through a real web browser for you. You describe a task in chat. The agent starts a Chromium browser in an isolated cloud container and does the work step by step: it opens pages, clicks buttons, fills forms, logs in, scrolls, extracts data and saves files. Every step shows up in the chat, and the browser itself is visible in the panel next to it.
It doesn't write a script for you to run. It uses the browser itself and hands you the data or files.
The tools the agent has
The agent works with a fixed set of tools. These are the ones you will see in a session:
Browser actions: click, type, scroll (including infinite scroll), select from dropdowns, hover, paste, fill date pickers, go back, go forward and refresh. It reads the page by taking a screenshot and listing the clickable elements, and it can search a page for text or elements.
Extraction: single values, lists of repeated items, HTML tables, or the page structure. When selectors fail it can fall back to reading a screenshot with AI vision. See Data Extraction.
Tabs: up to four tabs in one browser, so it can work two sites side by side. Tabs share cookies and logins.
Web search and fetch: a quick web search, or fetching a page as plain text, without disturbing the browser.
Terminal and files: shell commands and Bash, Python or Node scripts in the session's Linux container. Python has pandas, numpy and matplotlib. It can read and write files there and turn any file it creates into a download.
Data tools: filter, reshape and summarize extracted data with JavaScript, then save it as CSV, Excel (several sheets if needed) or JSON.
Your apps: native integration nodes such as Google Sheets, Gmail, Slack, Notion and Airtable (see Integrations), plus tools from MCP servers you connect.
Results: a results table you can sort and download as CSV, with any generated files attached.
Sub-agents for parallel work
For broad tasks the agent can split the work. It hands out up to four sub-tasks at once. Research sub-agents only have web search and page fetching. Browser sub-agents each get their own tab in the same browser, with the same cookies and logins, so one can work on one job board while another works on a second. By default two browser sub-agents run at a time and the rest wait their turn. Sub-agents cannot log in, buy or post on their own. The main agent keeps those steps.
It asks you instead of guessing
The agent pauses and asks you in four situations:
It needs an account. If a task needs Google Sheets, Slack or another integration you have not connected, it asks you to connect one or pick an existing one.
It needs to log in to a website. With a stored username and password, the agent types them into the login form without ever seeing them. If the site asks for a 2FA code and you have saved a TOTP secret, it generates the 6-digit code and types that too.
The action is risky. Before paying, placing an order, sending an email or message to a real person, posting in public, deleting data or submitting a legally binding form, it shows you exactly what it is about to do. You approve or deny. If you deny, or don't answer, it does not go ahead.
It is stuck. On a CAPTCHA it cannot solve, or a page it cannot get past, it offers you Take Control so you can sort it out in the live browser. reCAPTCHA v2 and hCaptcha can be solved automatically when CAPTCHA solving is turned on. Others, such as Cloudflare Turnstile, need you.
You stay in control while it runs
The agent shows its step plan and ticks off each step. You can send a correction at any time ("that's the wrong button, use the pagination at the bottom"). You can also take over the browser yourself and hand it back. The agent then carries on from wherever you left the page.
The agent also learns your preferences. If you say "always use DD/MM/YYYY" or "remember our CRM is the customers table", it saves that to your workspace's agent memory, and later sessions start out knowing it.
From one chat to a repeatable workflow
The agent records every step it takes. When a session finishes, Convert to Workflow turns that record into a workflow in the visual builder. It keeps the steps that worked and leaves out the failed attempts and inspection steps. You can then schedule it, add a webhook trigger, or edit it. Some things don't carry over. Agent to Workflow explains exactly what does.
You can also start agent sessions from your own code. The agent trigger API takes a prompt, authenticated with a scoped workspace API key, and returns the status and result. See the developer docs.
Limits worth knowing
Sites with strong bot protection can still block the browser. Autonoly does not offer proxy rotation or residential IPs.
Each step the agent takes uses credits. See pricing.
The agent works in its own cloud browser, not in your browser, so it starts logged out. You log in through stored credentials, or by taking control and logging in yourself.