JA EN
shiichan

Code Gets Written Faster Than It Can Be Reviewed — Cloudflare Unveils the Agent Development Lifecycle!

Cloudflare Blog ★★★★ View Markdown
#cloudflare#ai#devops#serverless

Hi, I'm Shii-chan! Today I want to share something pretty ambitious from Cloudflare. It's a proposal that looks head-on at how we need to rethink the way we build software now that AI agents can write code.

Cloudflare Blog blog.cloudflare.com

What was announced?

In a post on the Cloudflare Blog, Cloudflare introduced a new concept called the Agent Development Lifecycle (ADLC), along with a set of primitives that support it.

The post starts from a specific problem: agents keep getting faster at writing code, but teams can't review, deploy, and maintain that code at the same pace. So instead of just reusing the development flow built for human teams, Cloudflare says it's laying the groundwork for a "software factory" where agents themselves drive the process.

The story so far

The classic software development lifecycle (SDLC) has seven stages: Plan, Design, Implement, Test, Deploy, Maintain, and Retire. That framework has been around since 1975, so it's a well-established model.

Until now, Implement was the slowest and most expensive stage, so review and deployment processes were essentially designed around a human pace. With AI agents, implementation has become the fastest and cheapest stage instead. That flips the bottleneck onto review, deployment, and maintenance. The post points out this is already showing up in the real world: open source maintainers get overwhelmed by a flood of PRs and issues, and production engineers can't keep up with the sudden jump in delivery speed.

What changes

What Cloudflare calls a "software factory" is a system where, once you give it an input, agents autonomously build, improve, deploy, and take care of the software from there. Getting there means you can't just hand agents tools designed for humans — you have to redesign them specifically for agents. It's the same idea as an autonomous vehicle carrying sensors and communication systems a regular car doesn't have.

The post lists the properties this kind of infrastructure needs:

  • Programmatic (ClickOps-style clicking through dashboards doesn't work for agents)
  • Horizontally scalable (you need a production-like preview environment for every agent)
  • Reproducible (unit and integration tests alone aren't enough)
  • Real-time and push-based (humans watching a dashboard doesn't scale)
  • Atomic (each change can be tested, released, observed, and reverted independently)
  • Permissioned (a mechanism for escalating permissions as needed)
  • Self-improving (able to learn from experience)

In the end, this is a statement that Cloudflare wants to build the infrastructure for agents to ship code safely at production quality. The message is that humans get freed from the fine-grained work of supervising review and deployment, and can spend more time on the things only humans can do, like design and talking with customers.

Dive Deep

The first concrete primitive the post introduces is @cloudflare/ci. It's a CI tool built on top of Cloudflare Workflows, and it can chain multiple steps together, automatically retry failed tasks, and persist state for anywhere from minutes to weeks. Here's the code example from the post:

const deps: CiRunnerResult = await ci.runner({
  name: 'install',
  command: 'bun install --frozen-lockfile',
  cache: { inputs: ['package.json', 'bun.lock'] }
});

You can see it caching the dependency install step as it runs. A second example shows a workflow built with a framework called Flue, chaining data collection, agent deployment, running a review, and loading results, passing context along at each step.

The post also maps Cloudflare's capabilities onto each stage of the SDLC:

  • Plan / Design / Implement: a local dev environment built on Vite, Rolldown, and Oxc, plus Local Explorer, Local Traces, Remote Bindings, and Preview URLs
  • Test: Browser Run, a programmable browser that runs in the cloud, plus Vitest integration
  • Deploy: Flagship, which gives every change its own feature flag, plus Gradual Deployments
  • Maintain / Retire: Workers Logs, Agent Traces, the Cloudflare MCP Server, and Analytics Engine

Underneath all of that, the post names @cloudflare/ci, Cloudflare Workflows, and Artifacts as the three foundational primitives, meant to be a shared base that anyone can use, from startups to large platforms.

Wrap-up

Here's today's rundown:

  • Cloudflare introduced the Agent Development Lifecycle (ADLC) and the primitives that support it
  • The backdrop is a reversal: AI agents made implementation the fastest, cheapest stage, which turned review, deployment, and maintenance into the new bottleneck
  • Agent-ready infrastructure needs to be programmatic, horizontally scalable, reproducible, real-time and push-based, atomic, permissioned, and self-improving
  • The core building blocks are @cloudflare/ci, Cloudflare Workflows, and Artifacts, with each SDLC stage mapped to features like Local Explorer, Browser Run, Flagship, and Agent Traces

If you're building on Cloudflare's Workers stack, or you're on a platform team thinking about handing agents everything past implementation, this is worth a careful read!

Related posts