“engineers don't trust AI opinions. They just trust clear findings tied to standards that we … agreed upon. … it stopped sounding more like, maybe this is a problem and started sounding more like this violates a standard that we actually care about.”
Context. Describing early misses with Cloudflare's internal AI code reviewer, before the team tied its findings to machine-readable engineering standards.
Rajesh Bhatia, Senior Director of Engineering, Cloudflare. Leads the team that built and runs Cloudflare's internal AI engineering stack, so he is in a position to know its adoption and design; Cloudflare also sells the platform products it uses internally.
Cloudflare sells AI or agent tooling, so it has a reason to publish flattering numbers about AI. The passage stays in the archive; read it with that in mind.
How this was checked
Not yet checked by an independent second reader. The wording, speaker and date come from the source linked above; treat it as a lead until it is confirmed.
First reader (reader-web lane, 2026-11-01): two passes over the original audio, a whisper transcription of the full recording and a separate pass over the YouTube caption track, agree on the words quoted, except as noted. Not yet checked by a blind second reader. Cuts: 'have'/'ve' before 'agreed' (passes differ: 'we've' vs 'we have') and the sentences between. Both passes agree on the rest; caption pass filler 'uh' omitted. Punctuation only otherwise.
How to cite this passage
Rajesh Bhatia (Cloudflare), “From Assisted to Delegated: Cloudflare's AI Engineering Stack (interview, Background Agents Summit 2026)”, Ona (YouTube), 2026-05-06, at 26:40. https://www.youtube.com/watch?v=MbLdrAZFQRs Archived at https://the-ai-record.com/passages/cloudflare-engineers-dont-trust-ai-opinions-clear-findings-2026
Context. Asked which month or model made AI coding tools start to work at Cloudflare; he had said the company began piloting coding assistants in mid-2025.
Context. Answering how Cloudflare reached 93% of R&D using AI tools; he has just said adoption was slow at first because the models were not good enough.