It feels like we've moved past the point of prompt engineering being the de-facto bleeding edge of AI tooling and into system building.

That's had me thinking about how to identify and break down what I'm doing into the bits and pieces that go into making up a bigger system. Something like putting together a newsletter issue (or any content for that matter) doesn't tend to get very good results from a single prompt. But if you take that task and break it down into the steps that go into it and craft skills for each of those, you start to get better and better results.

That's a workflow I've been putting together to help me more efficiently write this very newsletter. I save content and notes throughout the week and have an agent organize and tag them so I can better group them into themes. Then I can either start writing this intro section, or if I'm struggling for a direction, have it prompt me with some ideas. Then have it proofread against my poor grammar skills.

Identifying all the items that go into a bigger task isn't easy. And designing systems that make it more efficient takes thought and skill.

An idea I've been working on is the ladder of complexity (working title, but it’s late and I gotta post.). The idea that you build up from prompt, to skill, to workflow, to agentic systems as you break work down into systems.

Prompts are your base for one-off help. You need answers now from general LLM knowledge. Skills are repeatable items with defined outcomes. Workflows are layers of skills in sequence to get a bigger task done. Agentic systems are used when tasks need dynamic decisions and tool use.

That last one takes infrastructure. It needs to be able to run without the human constantly in the loop guiding every action.

Which brings us to Cloudflare. They just completed their Agents Week and dropped all kinds of new features and tools for building an agentic web.

When you look at what they shipped through the building blocks lens, they're doing the same exercise but for infrastructure. Breaking down what agents need into granular, composable services. Not building "an agent platform." Building the primitives that agent platforms are made of.

Given the sheer number of drops from Cloudflare last week we'll shift the format a bit from our usual 5 links.

Agents have their own computers with Sandboxes GA
This is a full development environment for agents: terminal access via PTY and WebSocket, persistent code interpreters (variables survive between calls like a Jupyter notebook), background processes with live preview URLs, filesystem watching, and snapshots for fast session recovery.

Dynamic, identity-aware, and secure Sandbox auth
Outbound Workers act as programmable egress proxies for sandboxes. The big deal: agents never see credentials. Auth gets injected at the network layer, per-domain, per-sandbox. You can also dynamically change networking rules on the fly. Boot a sandbox, allow NPM and GitHub for dependency installation, then lock down all HTTP after setup. Zero-trust for agent compute.

Durable Object Facets: stateful code generated on-the-fly
Dynamic Workers can now instantiate Durable Objects with their own isolated SQLite databases. Durable Objects are pretty cool. It allows workers to have their own persistent storage.

Rearchitecting the Workflows control plane for the agentic era
Workflows got rebuilt for multi-step agent processes. Durable, retryable orchestration with state management. Think of it as the layer between your agent's decisions and the actual execution. When an agent needs to clone a repo, install dependencies, run tests, read failures, edit code, and run tests again, something needs to manage that sequence reliably.

Agents that remember: introducing Agent Memory
A managed service for persistent agent memory built on Durable Objects, Vectorize, and Workers AI. Each memory context gets its own isolated storage and vector search. Agents can recall what matters and forget what doesn't. If you're thinking about data feedback loops and agents that get smarter over time, this is a significant building block.

Cloudflare's AI Platform: an inference layer designed for agents
AI Gateway is becoming a unified inference layer. Call models from 14+ providers through one API with caching and observability built in. Plus a Workers AI binding for running third-party models and an expanded catalog including multimodal models. Cloudflare is positioning as the routing layer between your agent and every LLM.

Cloudflare AI Search
A managed search service that connects your data (websites, unstructured content) and automatically creates a continuously updating index you can query with natural language. RAG as a service, basically. Create search instances, upload files, and query across them with hybrid retrieval.

Scaling MCP adoption: enterprise reference architecture for MCP
Cloudflare's internal playbook for governing MCP using Access, AI Gateway, and MCP server portals. They also launched Code Mode to reduce token costs and new Gateway rules for detecting Shadow MCP. Yes, Shadow MCP is a thing now. Unauthorized MCP servers connecting to your infrastructure, treated the same way security teams treat shadow IT. If you're deploying MCP at work, this reference architecture is worth bookmarking.

Managed OAuth for Access: make internal apps agent-ready
Agents authenticating to internal apps through proper OAuth flows instead of shared API keys or service accounts. One of the hardest unsolved problems in agent infrastructure is identity and auth.

Introducing Cloudflare Mesh
Secure private network access for agents via Workers VPC. Gives agents scoped access to private databases and APIs without opening anything to the public internet or setting up manual tunnels.

Building a CLI for all of Cloudflare
A unified CLI called cf covering nearly 3,000 API operations, plus Local Explorer for debugging local data. Explicitly designed for both humans and AI agents to interact with the platform.

Cloudflare Email Service: ready for agents
Email as a first-class platform primitive in public beta. Agents can send and receive email programmatically. One less third-party service to wire up for agentic workflows.

Agent Readiness score
A score that measures how well your website supports AI agents. They explore emerging standards, share Radar data on current agent traffic, and detail how they made Cloudflare's own docs the most agent-friendly on the web.

Redirects for AI Training
robots.txt tells crawlers what not to index. But soft directives don't actually stop them from ingesting deprecated content. This lets you redirect verified AI crawlers to canonical pages with one toggle. No origin changes needed. A practical tool for anyone who cares about what version of their content is being trained on.

Thanks for reading, Jason