Cloudflare ran Agents Week from April 20 through 22. Roughly twenty-eight announcements over three days. I work there, and even I needed a second pass to absorb the scope.
Most of them are infrastructure plays for big customers. Four are the kind of thing I’d rebuild a personal agent stack around if I were starting from scratch today.
#1. Project Think (next-generation Agents SDK)
Cloudflare previewed a reimagined Agents SDK, framed as a “batteries-included platform for AI agents that think, act, and persist.” The relevant word is persist. The previous Agents SDK was good for stateless request-response patterns. Project Think is shaped for agents that hold state across days and weeks.
If you’re building an agent that should remember last Tuesday’s decision when it makes Friday’s, this is the substrate. I’m waiting on the GA release before I rebuild Hope’s runtime, but the preview docs already changed how I’m structuring her memory layer.
#2. The AI Platform inference layer
A unified inference layer that lets you call models from 14+ providers with a single Workers binding. OpenAI, Anthropic, Mistral, Together, Replicate, plus a growing list of hosted open models. Multimodal where the underlying model supports it.
This is the right shape for the model-pluralism conversation that’s been happening in the agent community for six months. You stop picking one model. You route to the right model for each call. Cheap for triage, frontier for reasoning, multimodal where the input demands it.
I’ve been doing this manually in OpenClaw with the OpenRouter provider. The Cloudflare AI Platform makes the same pattern native, with Cloudflare’s caching and observability layered in.
#3. Agent Memory Service
Persistent memory built specifically for agents. The framing in the announcement: “recall what matters, forget what doesn’t, and get smarter over time.”
I run Hope on file-based memory right now. AGENTS.md, SOUL.md, daily notes, a curated MEMORY.md. It works because she’s one agent on one machine. The minute you have multiple agents that need to share memory, or one agent that has to scale beyond what fits in markdown, you need a real service.
Agent Memory Service is the productized version of what I’d otherwise build with Supabase + pgvector + an MCP server. If you’re starting fresh, take the managed version.
#4. Managed OAuth for Access
Agents can now authenticate on behalf of users without insecure service accounts, via RFC 9728 compliance. This sounds dry. It’s the most important security announcement of the week.
Almost every “my agent acts on my behalf” pattern today involves either (a) a long-lived service token the agent holds forever, (b) the user logs in and the agent borrows the cookie, or (c) the agent literally types the user’s password somewhere it shouldn’t. None of those are good. Managed OAuth gives agents a sanctioned, scoped, revocable way to act as the user.
The tie-in to my BSides talk on credential takeover is direct: most account takeover attacks succeed because long-lived credentials are too easy to harvest and reuse. RFC 9728 narrows the blast radius.
#What I’m actually doing about it
Hope is staying on her current OpenClaw + claude-cli + Ollama setup for at least another month. She works. Don’t fix what works.
But the next agent I spin up (probably the Brand Strategist sub-agent for letsgochristo.com) will be on Project Think + Agent Memory Service + Managed OAuth from day one. That’s the post-Agents-Week architecture. The previous shape was already legacy by April 22.
Skipping the other twenty-four announcements isn’t a slight. They matter for enterprise. These four are the ones that change a one-person agent stack.
Let’s go!