Day 1 · Recorded 6 May 2026

Cloudflare's path from assisted coding to delegated engineering

If your AI rollout is stuck with a few early adopters, Cloudflare's story is the useful counterexample. Rajesh Bhatia walks through the platform primitives, identity controls, context systems, and code-review gates that helped Cloudflare move from assisted coding to delegated engineering across 93% of R&D.

Rajesh Bhatia, Senior Director of Engineering, Cloudflare

What's in this session

Enterprise AI adoption does not happen because a team buys better coding tools. At Cloudflare, the useful work came from building a platform layer around AI: context, trust, controls, visibility, and low-friction defaults that made the tools relevant to real work.

In this interview, Rajesh Bhatia walks through the stack behind that adoption: AI Gateway for routing, controls, and visibility; Access for identity; Backstage and Agents.md for service context; MCP infrastructure; one-command setup and shared config; and a multi-agent code reviewer tied to machine-readable engineering standards.

For developer productivity and platform leaders, the takeaway is a practical adoption playbook. The shift from assisted coding to delegated engineering depends on context systems, cost attribution, evals, objective review gates, and defaults that let engineers trust agents with more consequential work.

Inside the recording

  1. 00:00 Cloudflare's AI engineering stack in practice
    Rajesh explains the platform work behind AI use across thousands of Cloudflare employees.
  2. 03:00 AI Gateway, Access, Backstage, and Agents.md
    Cloudflare reuses its own primitives for routing, identity, context, and internal controls.
  3. 06:00 Getting to 93% R&D adoption
    Low-friction defaults, one-command setup, shared config, and real work drove the inflection.
  4. 17:30 Backstage as the system map for agents
    Ownership, dependencies, service boundaries, and conventions reduce how much models have to guess.
  5. 21:00 Measuring value beyond PR throughput
    Requests and tokens show usage, but feature velocity and safe change flow are the harder metric.
  6. 27:00 AI code review as an objective gate
    Engineering codex rules turn subjective model comments into findings tied to real standards.