Most of a workflow is a recipe. When a Stripe payment comes in, add a row to the revenue sheet, then post it in #sales. You know every step before it happens, so Lodol runs the same steps the same way every time.
Some jobs don't have a recipe. A line on your bank feed reads "SQ *BLUE BOTTLE OAKLAND CA", and someone has to work out who was paid and what for. The first search might settle it. Or it might lead to a second search, a look at the vendor's own website and a check of where they're based. A person decides what to look at next from what they just found, and until now, that was where a workflow stopped and someone's afternoon began.
The new Run an AI agent step is for that part. You give a model a task and a short list of tools, and it works out which to use, in what order, and when it's done. Then it hands back exactly the fields you asked for, and the rest of your workflow carries on as usual.
How you set it up
The step sits in the step picker's Advanced section, beside Code, HTTP and Workflows. It has a handful of fields:
- Model. Choose from Anthropic, OpenAI, Google Gemini, Mistral AI, DeepSeek, Groq and six more providers. The agent runs on your own account with that provider, connected once under Integrations. Lodol supplies the agent, not the model.
- Task. What to do on this run, written the way you'd brief a colleague, with the values it needs: "Find out who the vendor behind this bank line is: @line.description".
- Instructions. The standing rules for every run: whose books these are, what counts as found, what to leave out.
- Tools. What the agent may use while it works.
- Fields to return. The answer you want back, field by field.
- Files. PDFs or images for it to read.
Tools from your apps and your workflows
Any action from an app you've connected can be a tool: a web search, reading a page, looking up a contact in HubSpot. So can one of your own workflows. If you've already built a customer lookup, the agent can call it, fill in its inputs and read what it hands back. That's how you give an agent a fixed procedure for part of its job.
For any input you don't want the model to choose, fix its value. Say the agent may add a row to a spreadsheet: set the spreadsheet yourself, and it's filled in on every call and hidden from the agent. The agent decides what goes in the row, never which sheet it lands in.
Keep the list short. Two to four tools is usually right, because an agent with fewer choices makes fewer wrong ones.
A result the next step can use
Without Fields to return, the agent answers in text. With them, it has to fill in every field before it finishes, and later steps read each one by name. Describe each field, including the answers you'll accept:
| Field | Description |
|---|---|
| category | Exactly "urgent", "question" or "other" |
| reason | One sentence on why |
Then branch on it with an ordinary If: when the category is "urgent", alert whoever is on call. That's the dependable way to act on an AI's answer. You test a field the agent had to fill in, rather than hoping a particular word turns up somewhere in a paragraph.
It reads your files
Give the agent up to 10 PDFs or images at a time: a file you upload, a link, or an attachment an earlier step downloaded, such as Gmail's Download attachments. Reading files needs a model from Anthropic, OpenAI or Google Gemini. To pull the vendor, total and due date out of an invoice, give the agent the file and the fields to return, and no tools at all.
Every move on the record
An agent chooses its own path, so you need to see the path it took. Every run shows the model that answered, the tokens it used and every tool call it made: what it asked for, whether the call worked, and the start of what came back. When an agent reaches an odd answer, you can read exactly how it got there, the same way you'd read any other step in the run.
Built to finish the job
- A fallback model takes over if the first one fails to answer, during a provider's outage, say.
- A turn limit caps how many rounds the agent takes before it has to finish: 10 unless you change it, and at most 30. Each turn is one call to the model, so the limit also caps what one run can cost you.
- An optional memory lets runs that share a key, such as a Slack thread or a sender's email address, see what was asked and answered before. That's what a workflow that answers messages needs. Leave it empty for one-off jobs, so each run starts fresh.
When not to use an agent
When you already know the steps, build them as steps. Search, then read, then write is three ordinary steps, and ordinary steps run the same way every time at no extra cost. An agent calls a model on every turn of every run, and it may take a different route each time.
If all you need is one yes-or-no decision, such as whether an email is a sales lead, a condition set to Use judgment is simpler. It answers one question, says how sure it is, and costs nothing extra.
The agent is for the middle of a process, not the whole of it. The trigger, the checks before it and the steps after it still run exactly as you drew them. We think that's where AI belongs in work that matters, and our CEO wrote about why in Agents Are Having a Moment. Reliability Is Having a Reckoning.
Keep a person where it counts
An agent's answer is a suggestion until something acts on it. If the next step reaches a customer or moves money, put a Request approval step after the agent, so a person reads what it found and approves before the workflow goes on. Ask for the evidence as one of the fields to return, and the person approving has something to check.
Our bank-line templates are built around this kind of research. Verify the vendor behind a bank or card line gives an AI agent a Google search, a profile search and a page reader, and hands back who was paid, what they do, a suggested account, the evidence and the sources. Every fact has to come from a page it read or a result it saw, a guess is labeled as a guess, and a match with no source behind it is reported as not identified. Identify who was paid on every bank feed line runs it over each new line on a bank feed sheet.
What it costs
The agent runs on your own model account, so the provider bills you for what it uses, at its own rates; Lodol adds nothing. On Lodol's side, a run counts once however many steps it takes, agent included. Pricing has each plan's runs.
Give it the job with no recipe.
Connect one of the AI models it runs on, add Run an AI agent from the Advanced section of the step picker, and read how to set one up. Start free if you don't have a workspace yet.