Why turn an agent session into a workflow
An AI agent is good at working out how to do a task on a site it has not seen before. It is a poor fit for running the same task every morning. Each run costs more, takes longer, and can take a slightly different path. A workflow is the opposite: a fixed list of steps that runs the same way every time.
Autonoly connects the two. You figure out the task with the AI agent, then keep the result as a workflow. You don't have to rebuild it by hand node by node.
How the conversion works
During a session the agent records each action as a structured entry: the tool it used, its settings, whether it worked, and which data variable it produced. This record is kept separately from the chat itself, so it survives long conversations.
When the session finishes, the chat shows a Convert to Workflow button. Clicking it runs a converter on the server. The converter is ordinary code with no AI model involved, so the same session always produces the same workflow. Here is what it does:
- Keeps what worked. Failed steps are dropped. So are repeated attempts at the same click or URL, and back-to-back navigations within one site.
- Leaves out the investigation. Screenshots taken to look at the page, element listings, searches on the page, data previews, web searches and fetches, results tables and credential requests are all how the agent found its way. They are not steps the workflow needs.
- Drops dead ends. Extractions that came back mostly empty are removed, along with steps whose output nothing used.
- Maps agent tools to workflow nodes. Browser actions become Click, Type, Scroll, Select Dropdown, Hover, Fill Date and Navigate nodes. Extractions become Extract Data, Extract Table or Extract Child Collection. Data transforms become a Transform Data node. Saved files become Write CSV, Write Excel or Save As. Terminal commands become command, script or file nodes. Integration steps such as Google Sheets or Slack, and MCP tool calls, are kept as they are.
- Wires up the data. Where a step used another step's output (for example, rows extracted and then written to a sheet), the converter connects the two nodes so the data flows through on every run.
- Checks selectors against what is already known to work on that site (see Cross-Session Learning).
The result is saved as a draft workflow named after your prompt, and you get a notification with a link to it in the builder. The button then changes to View Workflow. If you convert the same session again, the previously converted workflow is replaced.
What does not carry over
Conversion copies steps, not judgment. Before you schedule a converted workflow, here is what to know:
JavaScript the agent ran on the page is mostly dropped. Simple scripts that press Enter or click an element are turned into Type and Click nodes. Anything else run that way is not converted.
Approvals do not carry over. If you approved a submit or a send during the session, the workflow does that step on every run without asking. Review those steps before scheduling.
Logins via credential autofill are not recorded. Neither is anything you did by hand during a takeover. If the site needs a login, add login steps in the builder, or use cookies.
Web search results are not re-run. If the agent searched the web to find a page, the workflow goes straight to the URL it found that time.
Values are fixed. URLs, search terms and text typed in the session are stored as they were. Edit them in the builder to change them.
Parallel browser sub-agents run one after another. Each one's steps replay in sequence, on its own tab.
SSH logins need a stored credential. If the agent connected to a server with credentials typed into the chat, the workflow asks you to select a stored SSH credential instead.
Very long sessions keep the most recent 500 recorded actions, so the earliest steps of a very long session may not be included.
At run time a workflow does not reason like the agent does. If the site changes a lot, steps can fail. Extraction steps retry with selectors already proven on that site, and Click and Type steps retry as they are. When a run does fail, fix the step in the builder or rerun the task with the agent and convert again.
After conversion: schedule and trigger it
A converted workflow is a normal workflow. In the builder you can test single nodes, add branches and error handling, and use version history to go back to an earlier version. You can then run it:
By hand from the builder.
On a schedule with a cron expression and a timezone (see Scheduled Execution).
From a webhook URL, so another app can start it (see Webhooks).
On an event that Autonoly checks at an interval you set: a new Gmail message, a new Google Sheets row, an RSS feed item, a change on a web page, a URL's status, or a polled HTTP endpoint.
Scheduled runs use the same kind of cloud browser as the agent, and you can watch them live and take control while they run.