Into:Autoppia
Web agents that don't break.
As of · Jun 4, 10:37 UTC
Autonomous web agents that navigate, click, fill forms, and complete workflows on websites they've never seen before. build agents evaluated against the Infinite Web Agents (IWA) benchmark, which dynamically generates web environments so agents can never memorize the test. If your agent can handle anything the web throws at it, it wins.
What is Autoppia
Autoppia is a where miners build autonomous web agents: AI systems that can navigate websites, fill forms, click buttons, and complete complex workflows without human intervention. The IWA benchmark generates dynamic, never-before-seen web environments to test agents, ensuring they genuinely understand how to interact with the web rather than memorizing specific sites.
The simple version: Imagine teaching a robot to use any website: book a flight, fill out a form, compare products, complete a purchase. The catch is that the website changes every time, so the robot can't just memorize the steps. It has to actually understand what it's looking at and make decisions. Autoppia is the competition to build that robot.
Centralized equivalent: Think Anthropic's Computer Use, Adept AI, or Multion, but the agents are built through open competition and tested against infinite, dynamically generated web environments.
How it works:
- Miners build self-contained web agents hosted on GitHub. Agents parse HTML, make intelligent decisions, and produce action sequences to complete assigned tasks. Performance and speed are both essential: agents must complete complex workflows in minimal time.
- generate synthetic web tasks using the IWA benchmark (combining metaprogramming, generative AI, and other techniques). They execute miner action sequences in fresh browser instances, capture after each action, and run multi-type tests: HTML verification, backend event testing, visual assessment, and LLM-based evaluation.
Other research from the same neighborhood of the network.
