Superpower Daily
OpenAI Loses Data Center Chief While Pursuing a Reported $600 B...Alibaba Raises HK$80B for Full-Stack AI as Investors Question t...LLMScholarBench Tests 22 Models and Finds an Accuracy-Represent...Meta Adds Voice-Controlled AI and Navigator Home Screen to Ques...Harvard Uses 7 AI Faculty Avatars in a $699 Founder Bootcamp, W...DeepSeek V4 Pro Lands on Fireworks With $2.50 CyberGym Solves,...OpenAI Loses Data Center Chief While Pursuing a Reported $600 B...Alibaba Raises HK$80B for Full-Stack AI as Investors Question t...LLMScholarBench Tests 22 Models and Finds an Accuracy-Represent...Meta Adds Voice-Controlled AI and Navigator Home Screen to Ques...Harvard Uses 7 AI Faculty Avatars in a $699 Founder Bootcamp, W...DeepSeek V4 Pro Lands on Fireworks With $2.50 CyberGym Solves,...
Productspublished

Ora Adopts Vercel’s Eve After Its Benchmark Found Fewer Steps and Lower Costs

The adoption turns Ora’s testing stack into its production stack, letting it trace its own eve-powered agents in the same environment used to inspect rival harnesses on customer websites.

By 3 min read
Ora Adopts Vercel’s Eve After Its Benchmark Found Fewer Steps and Lower Costs
Sign in to save and react

Listen to this story

The audio brief

About 1:33
0:001:33
Read transcript
Ora is now building its own AI agents on Vercel’s eve framework, turning the company’s testing environment into its production environment. That matters because Ora sends agents through real customer websites, and it needs to record not just whether a task failed, but where, at what cost, with what latency, and after which individual step. The choice followed a benchmark across hundreds of journeys and multiple domains. In the disclosed comparison, eve went up against Claude Code using the same Claude Fable 5 and Haiku 4.5 models, with both harnesses asked to integrate with a product. Ora reported that eve used 7% fewer steps, completed twice as many journeys directly on customers’ sites instead of falling back to web search, and found 9% more callable endpoints. After Vercel fixed a prompt-caching issue identified in one run, a later round showed roughly 15% lower total cost. There’s an important limit to those results: Ora says it tests six harnesses, but the published numbers compare only eve with Claude Code. The practical advantage is eve’s sandbox override. It lets Ora replace eve’s default execution environment with Ora’s own traced runtime, without building separate instrumentation. Ora already runs its frontend, backend, and agent runtime on Vercel, sharing deployment, logs, and authentication. So eve now sits on both sides: a harness Ora evaluates and the framework it uses internally. The constraint to watch is whether that single traced setup can support comparisons across all six harnesses without losing consistency.

Story brief

3 key points

Vercel’s eve framework is now Ora’s internal agent-building stack, extending a design partnership that also produced a limited benchmark against Claude Code. In Ora’s reported tests, eve used 7% fewer steps, completed twice as many journeys directly on customer sites, and found 9% more callable endpoints; a later round showed about 15% lower total cost after a prompt-caching fix. The important product implication is...

  1. 01

    Benchmark used Claude Fable 5 and Haiku 4.5 across hundreds of journeys on multiple domains.

  2. 02

    Eve’s sandbox override lets Ora substitute its traced environment and capture failures, costs, latency, and individual steps.

  3. 03

    Ora lists six tested harnesses, but the disclosed results compare eve only with Claude Code.

Ora, which sends AI agents through customer website journeys, now builds its own agents on Vercel’s eve framework. The practical consequence is that Ora can run those agents inside the same instrumented environment it uses to record failures, costs, latency and individual execution steps across its benchmark suite.

A benchmark informed the choice

Ora’s published comparison put eve against Claude Code across hundreds of real journeys on multiple domains. Both harnesses received the same task—integrating with a product—and used the same models, Claude Fable 5 and Haiku 4.5. Ora said eve took 7% fewer steps, completed twice as many tasks natively on the customer’s site rather than falling back to web search, and found 9% more endpoints it could call.

Those figures need a clear boundary. Ora works with Vercel Engineering as a design partner, and Vercel published the customer story that reports the results. The disclosed comparison is eve versus Claude Code, not a ranking across the six harnesses Ora says it benchmarks side by side.

Tracing is the production requirement

Ora treats an agent as two layers: a model that reasons and a harness that supplies tools and moves the model through a task. It lists Claude Code, ChatGPT, Gemini, Hermes, OpenClaw and eve among the harnesses it tests. Because each expects a different runtime and exposes steps differently, Ora runs a separate runtime for each and traces every step.

That setup created the decisive constraint for Ora’s own use of eve. Agent frameworks normally provide a sandbox: an isolated place where an agent runs tools and accesses files. Ora says eve’s sandbox override lets it replace that default environment with Ora’s traced environment, so an eve agent can be recorded without building separate instrumentation for it.

A single operating surface

Ora runs its front end, back end and agent runtime on Vercel with a shared deployment path, logs and authentication. That consolidation means eve is now present on both sides of the product: it is one of the harnesses Ora evaluates, and the framework Ora uses to build internal agents.

Ora’s larger premise remains demanding. The company estimates that 99% of the web is not ready for agents attempting to sign up, integrate with a product and pay for it. Its product is designed to show customers where those journeys break and what an agent attempted at the failed step; the eve adoption gives Ora’s own agents the same diagnostic treatment.

Sources

  1. vercel.comHow Ora benchmarks every major AI agent on Vercel | Customers | Vercel

Related coverage

Continue reading