How to Use Airtop Agent Builder: Build, Test and Deploy a Browser Agent

01

1. Start with one repeatable browser job

Choose a task with a stable goal and an output you can verify: collect rows from a portal, monitor pages, enrich a lead list, update records, run a recurring SEO check or move information between a browser-only system and a spreadsheet.

Avoid making your first agent a workflow where every run is fundamentally different. Agent Builder's main advantage is strongest when repetitive browser behavior can be compiled and only genuinely variable decisions require model reasoning.

02

2. Describe the outcome and let Airtop build it

Tell Agent Builder what should happen, which site or app is involved, what information should be extracted or changed and what the final result should look like. Airtop says the builder can ask clarifying questions and navigate a live browser during the build process.

Watch the builder's live execution and correct wrong actions before deployment. Airtop's architecture is designed to fold repeatable corrections into the compiled automation instead of making the model relearn the same browser step on every future run.

03

3. Handle authenticated sites carefully

Airtop supports agents that sign into websites and stores credentials in a secure vault. Use the narrowest account permissions practical for the workflow and avoid placing passwords or secrets directly inside prompt text or exported workflow descriptions.

For workflows that can edit records, send messages or submit forms, test against a low-risk account or sandbox where possible before granting production access.

04

4. Test the exact browser actions and outputs

A single successful run is not enough. Verify which pages were opened, which fields changed, what data was returned and how the agent behaves when expected elements are missing. Airtop provides logs and session replay to inspect runs.

Keep deterministic browser problems separate from reasoning problems. A selector or navigation issue should be corrected in the browser workflow; inconsistent judgment may require changing the model step or its instructions.

05

5. Deploy, schedule or trigger the agent

Airtop documents on-demand runs, cron-style schedules and webhook triggers. Choose the least aggressive cadence that meets the business need, then increase frequency only after the workflow produces stable outputs.

Before scaling, check your plan's deployed-agent and simultaneous-session limits. Airtop's current pricing page ranges from three simultaneous sessions on Free and Starter to 100 on Enterprise.

06

6. Add cost and time guardrails to AI-heavy steps

Airtop's AI APIs report credit usage in response metadata. Its documentation also supports costThresholdCredits and timeThresholdSeconds so applications can stop AI operations after configured thresholds are crossed.

07

7. Review repair drafts before scaling

If a deployed run breaks, Airtop can diagnose the failure, draft a repair and test it. The repair remains a draft until you choose to deploy it, so review the new behavior instead of assuming every automated fix is safe.

Once results are predictable, batch across lists of URLs or rows and increase concurrency gradually. Target-site rate limits, CAPTCHAs, authentication changes and runtime AI judgment can still affect throughput even when your Airtop plan permits more parallel sessions.