Justin and Avichal here. We’re building Linden (http://trylinden.com), a tool for maintaining browser automations at scale.

Avichal lived this problem at a dental insurance startup, where he spent time fixing brittle Selenium scripts that broke on minor website updates. He was pulling data from dozens of legacy portals and the constant maintenance was a drain on engineering time.

The new AI browser tools had their own issues. We found them to be mostly black boxes, lacking the observability and control needed for business-critical pipelines.

We're building Linden with a different approach: AI isn't responsible for the core execution; instead it simply creates and maintains Playwright scripts for you. This makes the entire workflow transparent and robust.

Here's how it works:

1. Record the First Run: To start, the task is performed just once by our AI agent (or manually). As the task runs, we record the sequence of raw, deterministic actions, like click(selector) and fill(selector, value).

2. Export a Workflow: This action trace is saved as a clean Playwright script. You also get the raw steps in JSON, giving you a fully transparent workflow you can read, edit, and run yourself.

3. Execute the Script: All future runs execute this script directly. It’s cheaper, faster, and more stable than having an AI agent re-evaluate the page from scratch every time.

4. Self-Heal on Failure: When a step in the workflow fails (e.g., a selector becomes stale), an AI agent is invoked with a narrow scope: to fix only that one broken step. It then patches the workflow, and the updated script is saved for all future runs.

We’re still super early and we’d love your feedback. Also interested to hear about how people are currently dealing with browser automations at any scale.