The archive · Product Ideas · Product decision · 2024
Skyvern lets vision LLMs read any website and automate it without XPath scripts
Skyvern pairs task-driven LLM agents with Playwright so browser workflows need no per-site selectors and survive every redesign.
Skyvern
What it had to solve
Classic browser automation and RPA meant custom scripts per website — DOM parsing and XPath interactions that broke whenever a layout changed and had to be rewritten for every new site. The open-source project Skyvern proposed replacing that with LLMs and computer vision: an agent that understands what it sees, the way a person does.
How it works
Before Skyvern, browser automation was a game of brittle specifics: teams wrote custom scripts that located buttons and forms by DOM structure and XPath, so every site redesign broke the robot and every new site meant new scripts. RPA vendors sold this as unavoidable. The Skyvern open-source project, launched in March 2024, treated it as an obsolete assumption.
The project's idea was to give the agent eyes. Borrowing the task-driven agent designs popularised by BabyAGI and AutoGPT, Skyvern added Playwright browser control and a swarm of agents that comprehend a page visually, plan the steps that complete a workflow, and execute them with clicks, keystrokes and data extraction. Because nothing depends on a pre-determined selector, the same workflow can run on sites the system has never seen and survive layout changes.
The Show HN on 14 March 2024 drew 422 points and 139 comments, many of them about the darker implications — automated spam, bot-versus-anti-bot escalation, and whether AI agents would soon dominate web traffic. The founders replied that they declined abusive use cases and deliberately kept the core open source so site owners could audit what the tool can do.
Why it lands
- Selector-free automation inverted the maintenance cost: instead of rewriting scripts after every redesign, the agent adapts at runtime because it reads the rendered page.
- Vision is the interface humans and agents share: mapping visual elements to actions works on government portals, insurance forms and stores the developer has never opened before.
- Open-sourcing the agent set a new default for the category and answered the trust question raised in the launch thread: site owners could read exactly what the tool was capable of.
- An API-first design meant the browser agent could plug into existing automation stacks as a service, rather than asking teams to abandon their tooling.
What it did
The Show HN drew 422 points and 139 comments, with the founders answering questions about abuse, pricing and openness in the thread. By 2026 the GitHub repository showed about 22.9k stars and 2.2k forks, and the project's page claimed state-of-the-art results (64.4% accuracy on WebBench) for browser agents.
What you can take
When automation breaks because the world changes, replace the fragile assumption — page structure — with a capability: an agent that reads the rendered interface works on layouts nobody coded for.
Since then
The launch thread anticipated the debates that followed: commenters worried Skyvern would accelerate spam and trigger an arms race between AI agents and bot detection, while the founders said they rejected abusive use cases and kept the code open so the web could defend itself. The repository grew into a Playwright-compatible SDK and no-code workflow builder with a managed cloud, and by 2026 it showed roughly 22.9k stars and 2.2k forks, with the project claiming state-of-the-art accuracy on browser-agent benchmarks.
Sources
spotted an error? The archive wants to know.
Your turn
You just read one. Describe the brief you are staring at, and see who has been given the same problem.
Free account · 3 free questions · no card