Cloudflare's Agent Development Lifecycle: The ADLC Is Now a Platform Bet

On August 4, day three of Cloudflare's Agents Week, the company made its biggest statement yet about what it thinks agents will do to software engineering: it announced the Agent Development Lifecycle, or ADLC, and shipped the first platform primitives for it. The umbrella post, written by Brendan Irvine-Broque, argues that the SDLC's assumptions break once agents write code faster than teams can review, deploy, and maintain it. The fix, per Cloudflare, is not fewer agents, it is a platform where agents own more of the lifecycle, not just the implementation step.

Three product launches landed under that umbrella on the same day: Cloudflare Agents, a dashboard plus agent tracing built on OpenTelemetry; @cloudflare/ci, a way to run CI/CD pipelines as Cloudflare Workflows; and OpenTelemetry traces in local development via Wrangler and the Cloudflare Vite plugin. Cloudflare also published two dogfooding posts: how it enforces engineering standards with AI, and how it built a software factory that drove Astro's GitHub issue count to zero.

What shipped#

Agent tracing with session replay. Cloudflare Agents is a new dashboard view plus an agent tracing system. Agents built with Think, Flue, or the AI SDK emit spans for agent invocations, model calls, tool executions, approval events, and supported subagent calls, layered on top of the existing Workers infrastructure traces. The dashboard shows two debugging views per agent session: a Messages tab that replays the recorded conversation (system prompt, user messages, thinking, tool calls with arguments and results, final response), and a Traces tab with an execution waterfall that ties agent operations to the D1, KV, and Durable Object calls they triggered. Payload recording is opt-in per harness via storeMessages and storeTools, which matters when traces contain secrets or personal data. Custom harnesses can use the Workers custom spans API and follow OpenTelemetry's Generative AI semantic conventions, and Cloudflare says OTLP-compliant frameworks will work without adapters soon. Traces export to any OTLP-compatible provider.

Tracing is free while in beta. From October 1, 2026 it inherits Workers Observability pricing: the Free tier gets 200,000 events per day with 3-day retention, and Paid gets 20 million events per month included with $0.60 per additional million and 7-day retention. Every span counts as an observability event.

CI/CD as Workflows. @cloudflare/ci reframes a pipeline as a Workflow: steps run in containers, results flow between steps, and the same primitives that orchestrate long-running processes now run builds, tests, and deploys. The example in the post chains bun install with caching, runs lint, test, typecheck, and build in parallel, then deploys with Wrangler using account credentials. Because Workflows can spawn agents and other Workflows dynamically, a pipeline step can do more than run a command, it can dispatch an agent to investigate a failure or reproduce a bug before deciding whether to block a merge.

Local traces. The Wrangler CLI and the Cloudflare Vite plugin now emit OpenTelemetry traces in local development, so the trace you see before deploying matches the one you get in production. For agent work this closes a specific gap: remote bindings already let a local agent hit production D1, KV, and Durable Objects, and now the local debugging view is the same shape as the production one.

Why it matters to developers#

The ADLC framing is the meaningful part, more than any single feature. Cloudflare's argument is that the SDLC is a model for humans coordinating on shared code, and it assumes human-paced review, human babysitting of deploys, and humans holding the pager. Each of those assumptions fails at agent throughput, and the industry response so far has been to keep the SDLC shape and bolt agents onto individual steps. Cloudflare wants the opposite: agents should own whole stages, and the platform should be designed for them.

The umbrella post lists seven requirements for a platform that can survive that: programmatic everything (no clickops), horizontally scalable previews, reproducible environments, real-time push-based events, atomic changes, granular permissions, and self-improvement. It is the clearest public articulation yet of what a serious software factory needs from its infrastructure, and it lines up with what we have seen in the harness ecosystem for months. Our take in "Why software factories fail without harness engineering" was that the factory concept dies on unobservable, unownable agent runs. Cloudflare is now shipping the observability layer as a product, which is the direction that argument pointed at.

Three developer-facing takeaways:

• Agent debugging finally has a first-party home. An agent can return HTTP 200 and still be broken: wrong tool choice, stale context, token-burning retry loops. Traditional APM shows the API call, not the reasoning that caused it. Agent-aware traces with session replay answer the questions that matter ("did the turn pause for approval?", "which subagent did the work?"), and being able to export to any OTLP backend means the data is not locked in. Our Copilot traces analysis made the same point about production-scale agent traces: raw infra telemetry is not enough, you need the agent operations layer on top.

• The open-source observability stack won the stack wars. Cloudflare's choice to build on OpenTelemetry semantic conventions, rather than invent a proprietary agent telemetry format, is notable. The trace in the announcement shows a Travel_Planner agent calling a GLM-4.7-Flash model, hitting D1 and KV through tool calls, all in one waterfall. Standardizing on OTel means the Agent Development Lifecycle pitch scales beyond Cloudflare's own harnesses, which matters because Flue and the AI SDK both support deployment elsewhere.

• CI/CD is being repriced as orchestration. @cloudflare/ci says the quiet part: a pipeline is just a workflow, and workflows can do more than run commands. Once a failing test can spawn an agent that reproduces the bug, the line between "CI" and "agent platform" dissolves. That is a direct competitor to the dedicated agent-CI products that have been appearing, and it is coming from a platform that already runs the underlying primitives.

The honest constraints#

The announcements are day-one versions. Agent tracing supports three harnesses at launch (Think, Flue, AI SDK), and everything else relies on custom spans or waits for the OTel API work inside Workers. Session replay is replay of recorded data, not re-execution, so if the harness did not record payloads, the Messages tab has nothing to show. Pricing only becomes concrete on October 1, when tracing joins Workers Observability billing; between now and then the free beta could change shape. And the ADLC remains an argument, not a feature: the platform primitives exist, but the seven requirements (permissioned escalation, self-improvement, atomicity) are still largely on the roadmap, not in the dashboard.

Continue Reading#

• Copilot agent traces at production scale - what a real platform learned instrumenting millions of agent runs
• Why software factories fail without harness engineering - the failure modes Cloudflare is trying to engineer around
• Flue as an agent harness layer - the harness behind one of the three tracing integrations
• Cloudflare Computer: agent runtime preview - where Cloudflare's sandboxed agent execution is heading
• Long-running agents need harnesses - why the multi-hour agent that owns a full task changes the platform requirements

Sources#

• The Agent Development Lifecycle has arrived on Cloudflare - Cloudflare Blog
• Introducing: Cloudflare Agents - Cloudflare Blog
• Run CI/CD for millions of repos - Cloudflare Blog
• Your agent can now debug Workers with local tracing - Cloudflare Blog
• Cloudflare Agents tracing documentation - developers.cloudflare.com
• OpenTelemetry Generative AI semantic conventions - opentelemetry.io