AgentWeb claims 47/47 success on live-site tasks at 1.9s median and $3 total

AgentWeb is pitched as web infrastructure enabling agents to operate on sites lacking APIs, including government portals, insurance directories, and signed-in services such as LinkedIn, Amazon, and Uber Eats. The team says it benchmarked 47 real-world tasks across 15 government portals, 7 insurance providers, and 25 signed-in sites, claiming 100% success, a 1.9-second median completion time, and $3 total cost. The authors claim it outperforms state-of-the-art computer-use models on speed, reliability, and cost. For sites without API specs, it captures button actions to build a typed action map, and patches in real time when backends change. Access is by request; the project was built over six months.

✨ AI summary of the linked source — not editorial content.