I feel very lucky to be sitting at a very exciting frontier at Cloudflare. It can be hard to discern noise and reality in Silicon Valley. But the work we're doing on agent payments is real, and we're building it together with the best partners in the industry.
We have so much work ahead of us, and I'm more excited than ever.
The logic that orchestrates your compute rarely shares the same resource requirements as the compute itself. Do you really need a giant box to do inference with a model provider? Or do you just care about the big box when you’re compiling or running tests? This pattern extends to many different use cases.
We’re working really hard behind the scenes to make durable objects and containers work seamlessly together as compute primitives. Colocating DOs and containers is just one example of this. Dehydrating and hydrating disk quickly, scaling up resources programmatically, and integrating well with all of the goodies Cloudflare’s developer platform has to offer will make agentic compute just make sense.
Computer is one of those ships that I am truly excited about - and while it looks like we just glued a Durable Object and a Container together, making this work is what we've been building towards for so long.
It lets your Agent run in an incredibly efficient, durable environment while delegating all of the heavy lifting to a container. If you've ever tried to separate your agent harness from your sandbox , you know how tricky it is to keep their states in sync especially when something crashes.
Computer lets you avoid the complexity, and why it works is the most interesting bit. We actually attach a private container to your Durable Object, allowing it to share the same underlying DO storage via a custom FUSE filesystem.
At first this looks a bit like network attached storage, and it is - but next up we're going to ensure that your Durable Objects and their container side cars are co-located on the machine, allowing shared access to be blazing fast.
Computer is one of those ships that I am truly excited about - and while it looks like we just glued a Durable Object and a Container together, making this work is what we've been building towards for so long.
It lets your Agent run in an incredibly efficient, durable environment while delegating all of the heavy lifting to a container. If you've ever tried to separate your agent harness from your sandbox , you know how tricky it is to keep their states in sync especially when something crashes.
Computer lets you avoid the complexity, and why it works is the most interesting bit. We actually attach a private container to your Durable Object, allowing it to share the same underlying DO storage via a custom FUSE filesystem.
At first this looks a bit like network attached storage, and it is - but next up we're going to ensure that your Durable Objects and their container side cars are co-located on the machine, allowing shared access to be blazing fast.
@samgoodwin89@hackerdocc@alchemy_run There are ways with our current architecture that we can provide much better performance than what you are observing. We don’t typically offer these enhancements by default for various reasons, but are working on ways to bring more consistent performance to all
We are launching Workers Cache, a regionally tiered cache that sits directly in front of your Worker entrypoints. Infinitely composable, configured via standard HTTP headers. https://t.co/eBbxHIgUBA
@_dylanga I have a skill for this that persists state to an .md for easy rehydration but recently compaction has been pretty solid for me. Don't really notice any session degradation
Excited to launch Claude Managed Agents on Cloudflare today!
- Run sandboxes as microVMs or even lighter-weight isolates on CF
- Zero-trust creds injection, custom egress proxies, better observability, private services via VPC
- Agent Email + Browser Run + easy custom tool calls
Cloudflare Sandboxes now do secure credential injection.
We use Workers as a trusted proxy, which make it really easy to customize egress policies (globally or per-instance) and update them dynamically. Let's take a look...
Inject a token without the agent ever seeing it:
Sandboxes and Containers now support secure credential injection and dynamic egress policies. Upgrade to [email protected] or [email protected] to keep secrets out of your agents and control outbound traffic.
https://t.co/bxl3RBvRVp
@Stone_Tao@jsuarez I’ve been writing code exclusively with models for a few months now. Simple and complex changes alike. If you’re ever in a spot where the model is spitting out thousands of lines, I’m really not sure what you’re doing. You can break any problem down and ensure quality that way
Learn more about udpgrm, a lightweight daemon for graceful restarts of UDP servers. It leverages SO_REUSEPORT and eBPF to route new and existing flows to the correct server instance. https://t.co/vcWQQZeQit #DeepDive
Ran into an issue with a tokio::select! today where the the 2nd branch polled a oneshot receiver and pattern matched it on Ok(), then outside the select we called now_or_never() on the receiver just in case. The sender was dropping though so the now_or_never() would panic