Welcome to day 2 of Cloudflare's Agents Week, August edition! Today is all about the ADLC (Agent Development Lifecycle): 🔁 Cloudflare CI/CD 💳 Cloudflare Wallet 🤖 Agent Tracing for any AI Agent 📈 Otel Traces in wrangler dev 🧬 ADLC on Cloudflare 🏭 Building a Software Factory ✅ Enforcing AI Standards at Cloudflare Let's dive right into it! 🔁 Cloudflare CI/CD 🔁 Available today, Cloudflare's new CI package allows you (and your customers!) to define and run your own CI/CD pipelines, backed by Cloudflare Workflows, Artifacts and Sandboxes It's now possible to trigger a CI/CD pipeline on push to an Artifacts repo, in the same way you'd expect a push to GitHub to trigger a GitHub Action. The best part? Cloudflare Workflows, and therefore Cloudflare CI, are defined with TypeScript. This gives you near unlimited flexibility, and simplicity, versus wrangling with YAML. The features you'd expect exist today, including the ability to lint, cache dependencies, execute unit tests and of course deploy. 💳 Cloudflare Wallet 💳 In a world where AI is automating a growing list of tasks, there is often one blocker in the way: paying for goods and services. Humans still need to control that spend though, and that's where Cloudflare Wallet comes in. It will allow you to create a wallet, add funds, and distribute those finds for AI agents to spend. Each wallet has a stable id, allowing merchants to routinely charge that wallet. Naturally, there are controls in place. Your funds are stored in your account-owned wallet, from there you can create virtual wallets that can be assigned to AI agents - with the ability to set permissions and spending caps. Similarly, companies could create virtual wallets for all their employees to use. There's a lot of detail and depth to the blog post, so I'd recommend giving it a read - I think this will be the future of how transactions are handled on the web one day. 🤖 Agent Tracing for any AI Agent 🤖 One of the best uses of AI in software engineering is for debugging. Whether it be a bug, or a performance issue, AI is incredibly good at looking into these types of issues and providing plausible remediations. To maximize this potential, we need to be able to give AI agents insights into the problem. One of the ways to do this is with traces, showing the AI agent the exact order of execution for each AI agent session. It shows how long each span took, along with information like token usage, tools called and more. There's even session replay, allowing you to see the conversation as it played out. Available today, there's out of the box support for agent tracing for Think, Flue and the AI SDK. Any other frameworks can be instrumented too using the custom spans API, backed by Otel. That means you can instrument any AI agent, whether it be a Cloudflare-backed framework, or a third-party framework - all will be visible in the Cloudflare dash. In the near future, Cloudflare is working to support OpenTelemetry API directly in Workers - meaning no more custom adapters will be necessary. 📈 Otel Traces in wrangler dev 📈 If you love traces, today sure is your lucky day as there's yet more traces available with Cloudflare - this time in local development when using wrangler dev or vite dev. When Cloudflare's tooling detects an agent session, it points the agent to the Local Explorer API, a local debugging API where it can query those traces. This is available to humans too, of course, and there's no manual steps or dependenices to install. Agents can query a locally-exposed endpoint to retrieve traces, and humans can access the traces in their browser. Install the latest version of Wrangler and check it out for yourself! 🧬 ADLC on Cloudflare 🧬 Today's Agents Week theme is the Agent Development Lifecycle, so it's only appropriate there's an entire blog dedicated to it. With the huge shift in software engineering from implementation (writing the code) being the bottleneck, to review, deployment and other stages becoming bottlenecks, we need new approaches and tools to deal with this dilemma. All of the changes today aim to help streamline parts of the ADLC, with a lot more work to come, but there's already so many tools to empower agents on Cloudflare's developer platform: - Everything runs in local dev. - There are local traces as of today. - Remote bindings to reproduce production behavior - Preview URLs for every change - Browser run to give your agent a browser to test changes - Flagship and Gradual Deployments to control changes - Logs, traces, MCP servers and more to observe production behavior + address bugs There is truly no better place for your agents to build software than Cloudflare, and we're only just getting started. 🏭 Building a Software Factory for Astro 🏭 There's lots of talk of software factories, which is the ideal that you can assemble AI agents to produce working software (mostly) autonomously. Every production system produces exceptions and bugs, so a classic use case would be triaging and fixing those using a software factory. It sounds simple, but the actual results and impact vary greatly. This blog post examines how the Astro team within Cloudflare implemented a software factory, backed by Flue agents - an open-source agent framework. It's proved successful, too. Open issues are down from 200 to 30, with the number expected to hit 0 in the near future - the first time this will have happened for this project. If you want to read into how to build a successful software factory, this is the post for you! ✅ Enforcing AI Standards at Cloudflare ✅ I've previously shared how much I value the internal AI reviewer at Cloudflare. It's really good at catching issues, raising code quality concerns and other things you'd expect from a review. Unlike AI review tools that I've used in the past, where 90% of what is raised is noise, Cloudflare's AI reviewer is the opposite: 90% of it is typically useful. One of the reasons it's so helpful, as is the case with most AI, is what information it uses. Cloudflare has an internal codex, documenting the rules and principles we need to adhere to. There are a growing number of AI tools around standards, including SPEC reviewers and incident reviewers. In my opinion, a lot of AI is too generic. What works for one company isn't going to work for another. The value in AI is making it knowledgeable, and therefore valuable, for each use case. That's exactly why I believe the Cloudflare tools are as useful as they are. If you're looking to build your own internal AI tools, or just AI tools in general, this is a great place to start to get a guide of where to start, and the pitfalls to avoid. That brings us to the end of day 2, what was your favourite announcement of the day? Or what questions do you have on anything announced today?