Ora runs AI agents against live websites to measure how ready they are for autonomous tasks: sign up, integrate, pay. By Ora's own count, 99% of the web fails this test. The platform, built entirely on Vercel including the agent runtime, benchmarks six major harnesses side by side: Claude Code, ChatGPT, Gemini, Hermes, OpenClaw, and Vercel's own eve. Every run records cost, latency, and step count. Every failure shows exactly where the agent stalled and what it tried.
When Vercel launched eve, Ora gave it no special treatment. Eve ran the same benchmark as every other harness, against hundreds of real journeys across multiple domains, using Claude Fable 5 and Haiku 4.5. Three numbers came out: 7% fewer steps to reach the goal, 2x more tasks completed on the customer's native site without falling back to web search, and 9% more valid endpoints discovered. One benchmark run also surfaced a prompt-caching bug. The eve team shipped a fix, and the next round measured roughly 15% lower total cost. After those results, Ora built its own product on eve.
The full piece is worth reading for two specific details: how Ora's sandbox override lets eve agents run inside Ora's instrumented environment without new infrastructure, and how a 16-person team shipping hundreds of commits a day offloads day-to-day infrastructure work to coding agents built on the same stack. Ora is now splitting its platform into microservices, all on Vercel, with internal eve-based agents running as one more service in that architecture.
[READ ORIGINAL →]