Phase 1: Build

The Build phase is a conversation. You open a live browser, tell the agent what you want done the way you'd brief a colleague, and watch it work. The session you produce here (every page, every step, every note) is the raw material the compiler turns into a workflow, so a good demonstration is the single biggest factor in how good the compiled workflow is.

Launch a session

Click New Workflow on the Workflows page. The workbench opens and a dedicated cloud browser spins up for you in seconds: no local setup, no extensions. The browser fills the left side of the screen; the Builder chat sits on the right.

Talking to the agent

Plain English is the whole interface. Here is what a typical exchange looks like, from the Book Price Lookup example:

You: Go to books.toscrape.com, open the Travel category, and read out the title and price of the first three books.

Agent: I'll navigate to the site now. (browser steps: opens the homepage, clicks Travel in the sidebar, reads the listing) Here are the first three books in the Travel category, with prices in pounds, formatted as a table.

The agent narrates as it goes, and its browser actions collapse into a step list ("5 browser steps") you can expand on any turn to see exactly what happened.

One task or a step-by-step brief

You can hand over one complete instruction as above, or drive step by step and give the next instruction as each one finishes. Step-by-step briefs work well for longer processes:

1. Sign in to the Limit portal.
2. Open the Orders section and search for order 88-4417.
3. Read the fulfilment status and the shipping ETA.
4. If the status is "exception", open the exception details and summarize them.

The agent works through the list in order and reports as it completes each item. Numbered briefs also compile well, because each numbered step tends to become a clear step in the workflow.

A few controls worth knowing while it works:

• You can type mid-run. New messages reach the agent while it's working. Use them to correct course, add detail, or change the plan.
• The red Stop button aborts whatever the agent is currently doing, so you can redirect it immediately.
• You can ask questions. "What does the page say under shipping?" is a valid instruction; the agent reads and answers without changing anything.

Attachments and voice

Each message can carry up to 5 files (spreadsheets, PDFs, images) via the attach button or a straight clipboard paste. The agent reads them the same way it reads web pages, so "fill this form using the attached spreadsheet" just works. Prefer speaking? Voice input transcribes your instruction into the composer.

What the agent can do

The Builder agent operates the browser the way a person would, plus a few things a person can't:

When the agent asks for help

The agent doesn't guess when it matters. Four kinds of prompts can appear in chat:

Detail: A free-form question

"Which shipping option should I pick?" Answer in chat, or press Skip to let the agent use its judgment and move on.

Detail: Save website credentials

When a site needs a login, the agent raises a secure credential form. What you enter is stored in Passwords, reused by this workflow on every future run, and never echoed back into the chat.

Detail: Select the incoming message

For email-driven tasks, the agent lists recent messages from your team inbox and asks you to pick the example to work from. This is how you author inbox-driven workflows without ever opening webmail: the message you pick becomes the worked example, and future runs handle whichever message triggered them.

Detail: Action needed in browser

Some steps are yours to do: an SSO screen, a one-off confirmation. The agent hands you the live browser; do the step yourself and press Continue to hand control back.

"Noted for compile" cards

When you tell the agent something about future runs, it records the note and shows a green Noted for compile card in chat. These notes are how the compiler knows to create inputs, conditions, and OTP steps instead of hard-coding what it saw once. Two examples of what this looks like in practice:

• You say: "The order number will be different every time." A green card appears confirming the note, and at compile time the order number becomes a typed launch input (something like order_number) instead of the literal value you demonstrated with.
• You say: "Sometimes the portal emails a verification code at this step." The note compiles into an Email/OTP polling step: at run time the workflow watches your team inbox for the code and enters it on its own.

The same mechanism covers branches ("if the order is already shipped, skip the fulfilment step") which compile into Condition blocks. If you forget to mention something, don't worry: all of this can be added later in the Edit phase.

Workflow Settings

The gear icon opens settings that apply to authoring and to every future launch of this workflow.

Agent model

Changing the model mid-session takes effect immediately, with no browser restart. The tier you pick here is also the tier every future launch runs on.

Everything else

• Cache Persistence. Reuse this workflow's browser profile so login cookies and localStorage carry across launches. Fewer logins, faster runs.
• Extra Stealth. Launch browsers with managed stealth and CAPTCHA solving for bot-sensitive sites.
• Network RE Mode. Network activity is always recorded; this toggle lets the compiler use those captured calls, so repetitive UI work can compile into faster API steps instead of click-by-click browsing.
• Proxy. Choose a proxy type and country when a site expects traffic from a particular region.

Save: the session becomes a workflow

Click Save and the compiler reads the entire session (every step, every page, every green note) and builds the workflow graph. Compile progress streams live in the editor rail while the canvas grows block by block, so you can watch the workflow take shape. For the Book Price Lookup session above, the result is three blocks: a Workflow Input, a Parser that extracts num_books from the launch request, and the agent block that does the scraping.

Tips for a demonstration that compiles well

• Complete the task once, end to end. The compiler builds from what actually happened, not what you intended.
• Use realistic data. A real order number teaches the shape of the task better than "test123".
• Say what varies. "This date will be different every run" is one sentence for you and one input block for the compiler.
• Let logins go through the credential prompt so they land in Passwords and future runs can sign in on their own.

Saving lands you on the Editor tab with the fresh draft loaded, which is exactly where Phase 2: Edit picks up.