Cloudflare Ships Its Internal AI Stack and the Code Review Bots Are Already Winning

Cloudflare just shipped the AI engineering stack it built for its own R&D organization. The stat that matters: 93% of Cloudflare engineers used AI coding tools powered by this infrastructure in the last 30 days.

That's not a demo. That's production usage at a company with thousands of engineers shipping code that handles a meaningful chunk of the internet's traffic.

Seven agents per merge request

When a Cloudflare engineer opens a merge request now, it gets reviewed by up to seven specialized AI agents before a human sees it. Security, performance, code quality, documentation, release management, compliance—each has its own reviewer with a focused prompt.

The architecture avoids the "one gigantic prompt" antipattern. Instead of asking a single model to do everything, they spin up a coordinated set of smaller agents. Each agent has a narrow job and runs in parallel. The system doesn't try to replace human reviewers; it catches the obvious stuff so engineers can focus on the architectural decisions and edge cases that actually need discussion.

Cloudflare calls the orchestrator the "Review Plugin" inside their OpenCode process. It's a runtime plugin that coordinates the agents, merges their feedback, and posts consolidated comments back to the merge request. The agents run on the same infrastructure Cloudflare is now selling to customers.

Git-compatible storage built for agents

The other big piece from Agents Week is Artifacts: versioned storage that speaks Git. It's built so agents, developers, and automations can all treat code and data the same way—version control as a first-class primitive, not an afterthought.

Traditional agent frameworks often dump output into blob storage or databases without proper versioning. Artifacts gives agents a home for code and data that already integrates with the tooling engineers actually use. If an agent generates a config file or rewrites a module, that artifact lives in a Git-compatible namespace. You can diff it, roll it back, or merge it into another branch using the same commands you'd use for human-authored code.

Cloudflare also shipped secure private networking for agents, users, and automations. The pitch is that agents shouldn't live in a separate security perimeter—they need the same zero-trust controls as any other workload.

Parquet, Iceberg, and the logs no one wanted to store

Buried in the changelog is a quieter but equally useful update: Logpush now writes directly to Pipelines, which can store logs in R2 as Parquet files or Apache Iceberg tables. JSON logs are easy to generate and easy to read, but they're expensive to store and slow to query at scale.

Parquet is columnar, compresses well, and makes analytical queries orders of magnitude faster when you're dealing with petabytes of logs. Iceberg adds schema evolution and ACID transactions on top of that. If you're trying to debug an incident or build dashboards from raw logs, this matters a lot.

The analytics changelog also added Firewall for AI fields to the Firewall Events dataset: prompt injection scores, PII categories, token counts, and unsafe topic categories. Those fields are now queryable alongside the rest of your firewall telemetry.

The platform is the product

Cloudflare's play here is straightforward: they built this stack to solve their own problems, and now they're selling it. The code review agents, the Git-compatible storage, the private networking—all of it runs on the same infrastructure customers can provision.

That's a different pitch than "we built a general-purpose agent framework and think you should too." It's closer to "we run this in production every day, and here's the API."

The obvious question is whether the abstractions Cloudflare needed will map cleanly to what other engineering orgs need. Versioned storage and zero-trust networking are table stakes. The code review orchestration might be more opinionated—some teams will want to plug in different models or add custom agents, and it's not yet clear how much flexibility the platform offers.

But the usage numbers are real. 93% adoption in 30 days inside a large engineering org is not a pilot project. It's a signal that the stack works well enough to be boring infrastructure.

Elsewhere this week

OpenAI had a brief outage caused by an internal telemetry rollout. Not a security incident, not a bad deploy—just a config change that caused API and ChatGPT downtime.

Revolut announced its IPO is two years out and will be in the US, not the UK. The company also filed for a US bank charter, which makes the timeline more credible than the usual fintech IPO rumors.

Australia's ASIC joined the global regulatory pile-on around Anthropic's Mythos model. The FSB, Bank of England, and now ASIC are all monitoring a model that found zero-days in every major OS and browser. The capability is real; the question is what happens if it leaks.

The Cloudflare stack is the more interesting story this week. When a company ships the tools it uses internally and those tools are already at 93% adoption, that's worth paying attention to.