A benchmark lab that turns agent failures into fixes
Ora’s pitch is blunt: it sends AI agents onto live websites with real instructions — sign up for a product, integrate with it, pay for it — and watches them fail. By the company’s estimate, 99% of the web isn’t agent-ready. The platform’s job is to show customers exactly where and why an agent stumbles, and what they can change to fix it.
Co-founder Assaf Elovic came to this problem after selling Tavily, a search engine built for AI agents, to Nebius earlier this year. Search solved half the equation: an agent could find a product. But finding it isn’t the same as using it. With co-founder Liad Yosef, Elovic built Ora to close that gap by measuring agent readiness and repairing the weak spots. Today, Ora runs journeys from journey.ora.ai, recording the cost, latency, and step-by-step path an agent takes to finish a task. The whole platform, including the agent runtime, runs on Vercel.
Every agent gets its own runtime
An agent is really two pieces: the model that reasons and the harness — the software wrapping that model with tools and driving it forward. Ora’s test lineup covers the agents its customers actually use: Claude Code, ChatGPT, Gemini, Hermes, OpenClaw, and Vercel’s own framework, eve. Each one gets set loose on a customer’s website and is tracked through common workflows.
The hard part is that no two harnesses want the same infrastructure. Every one expects its own environment and exposes its steps differently. So Ora runs a separate runtime for each harness and traces every action. Engineering lead Ido Finder calls that side-by-side coverage one of the most valuable things Ora brings to customers. When an agent gets stuck in a signup flow, the customer sees which step it reached and what it tried. Without the trace, all you get is a score with nothing behind it.
Ora built this entire testing system on Vercel — front end, back end, and agent runtime share the same deployment path, logs, and authentication. The runtime isn’t separate infrastructure to operate; it lives where the rest of the product lives.
eve got no special treatment in the benchmark
When Vercel launched eve, Ora didn’t give it a pass. eve went through the same benchmark, under the same conditions, as every other harness in the lineup. Ora works with Vercel Engineering as a design partner, and Finder gave the eve team direct access to the platform so they could dig into the results themselves.
The first test pitted eve against Claude Code across hundreds of real journeys on multiple domains. Both harnesses ran identical models — Claude Fable 5 and Haiku 4.5 — and every run got the same job: integrate with a product. Ora published three results from the comparison:
7% fewer steps to reach the goal
2x native success: twice as many tasks finished on the customer’s own site instead of falling back to web search
9% more valid endpoints: more of the endpoints the agent discovered were ones it could actually call
The benchmarking flowed back into eve’s development. One run exposed a prompt-caching issue, the eve team pushed a fix, and Ora’s next round of results measured roughly 15% lower total cost.
“You don’t need to configure much on eve, it works out of the box, and parity with other harnesses is amazing to see.”
Ido Finder AI Lead @ Ora
For a company whose whole business is benchmarking every major harness, choosing a framework isn’t casual. After those results, Ora builds on eve.
Because eve follows the Next.js paradigm, there was little to configure, and tools, skills, and connectors take very little code. The feature that sealed it was the sandbox override. An agent framework like eve ships with its own sandbox — the isolated environment where the agent executes, runs tools, and touches files. That’s a solid default for most teams, giving safe execution for free. But an agent running in the framework’s own sandbox sits outside the instrumented environment where Ora traces every step. The override lets Ora swap that environment in, so eve agents get recorded like every other harness without building anything new.
Journey.ora.ai now has eve on both sides: it’s both one of the harnesses being tested and the framework the whole thing runs on.
“I’ve been building agents since the technology first became available, and eve was the easiest setup I’ve experienced.”
Ido Finder AI Lead @ Ora
16 engineers, hundreds of commits a day
The engineering team is just 16 people, yet they ship hundreds of commits a day — because the day-to-day infrastructure work belongs to their coding agents. A stack that lives in one place lets those agents operate it end to end.
“Everything is consolidated into the same infrastructure, which makes our coding agents much more efficient. I don’t even log into the Vercel UI. I ask my coding agents to run on the logs, debug everything, deploy, change environment variables." ”
Ido Finder AI Lead @ Ora
Finder puts the time saved at a few hours a week at least. Elovic credits a similar amount to how well coding agents build with Vercel’s libraries.
Scaling out without leaving the platform
Ora is adding more products, and the architecture is growing to match. The team is splitting its platform into microservices, all of them on Vercel. New services deploy to the same infrastructure and talk to each other with no extra configuration. The internal agents built on eve will run as one more service in that mix.
By Ora’s own measurement, 99% of the web still can’t handle an agent that shows up to sign up, integrate, and pay. Ora’s goal is to close that gap — and to measure its own progress while doing it.
