What happened
Ora, a startup founded by Assaf Elovic and Liad Yosef (previously behind Tavily, which was acquired by Nebius earlier this year), has built a benchmarking platform that tests whether AI agents can actually complete tasks on live websites — signing up for an account, integrating an API, or making a purchase — rather than answering benchmark questions in a lab. The company estimates 99% of the web still isn't reliably ready for agents to operate on. Ora's own front end, back end, and agent runtimes run entirely on Vercel, giving its 16-person engineering team a single deployment path, shared logs, and consistent authentication across every agent harness it tests, including Claude Code, ChatGPT, Gemini, Hermes, OpenClaw, and Vercel's own eve framework. Separately, GitHub published its post-incident review of the August 17 outage that knocked out github.com, authentication, Actions, and Copilot for nearly eight hours, citing a data-center component that failed to scale under record traffic.
Why it matters for your business
For any small business relying on third-party AI agents to book appointments, fill out forms, or complete checkouts on its behalf, Ora's finding is a pointed one: most sites aren't built to be operated by an agent, and that gap shows up as abandoned signups and broken automations before anyone notices. The GitHub outage is a companion lesson — even mature platforms hit capacity walls when usage spikes faster than infrastructure scales, and a retry storm from automated tooling can make an outage worse before it gets better. Businesses that lean on always-on integrations should ask their vendors how they test for both AI-agent compatibility and graceful failure under load.
What to watch next
Ora plans to keep publishing head-to-head agent benchmarks as new models ship, which should give buyers a clearer signal than vendor marketing claims. GitHub says it has already added millions of CPU cores and shifted more load onto Azure infrastructure, with a broader architecture rollout for large-repository operations still in progress — worth watching for whether it prevents a third August-scale incident.