Senior Principal Systems Engineer @cloudflare ETI
prev. @opstrace (co-founder/CEO, acq. by @gitlab), @redhat, @mesosphere, @cloudflare. @docker (co-founder)
spent the evening building a cardputer client for codex
can send tasks, watch edits come in, approve changes
cardputer → cloudflare sandbox → codex
kinda ridiculous, I know
Today, we share a breakthrough on the planar unit distance problem, a famous open question first posed by Paul Erdős in 1946.
For nearly 80 years, mathematicians believed the best possible solutions looked roughly like square grids.
An OpenAI model has now disproved that belief, discovering an entirely new family of constructions that performs better.
This marks the first time AI has autonomously solved a prominent open problem central to a field of mathematics.
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
you can build a Python agent that accepts a coding task, executes it inside a Cloudflare Sandbox, and copies the output files to your local machine
@OpenAIDevs x @CloudflareDev
Check out our guide here: https://t.co/qsUGwFce6z
One point of confusion I'm seeing: Lots of people assume you're supposed to run the AI agent harness itself inside a sandbox, and wonder how to do that with Dynamic Workers.
The harness does *not* need to be in the sandbox. It can be a regular old Worker (probably a Durable Object, probably using Cloudflare Agents SDK[0]). Remember that an LLM can't manipulate the world at all unless you give it tools. So if you are careful about what tools you give it, it is inherently sandboxed.
What we are proposing here is that you basically give the agent two tools: One to explore the APIs it has available to it (returning TypeScript type definitions), and one to execute code against those APIs (input is JavaScript). The latter tool runs the code in a Dynamic Isolate.
You can't quite just run Claude Code or Codex in there, since they aren't designed to run in Workers, at least today. But @southpolesteve managed to get OpenCode (a similar harness) running 100% in Workers[1]. It's also not as hard as you might think to write your own harness -- I've done it.
That said, if your agent's task specifically involves using a bunch of Linux programs, it may still be best to use containers for that. For instance, if your agent is doing software engineering -- checking out a git repo, running a compiler, etc. -- it may be impractical to make that all work in Workers. Try our Sandbox SDK[2] for that.
But if your agent is planning a family vacation, it probably doesn't need to run rustc. It probably just needs to interact with a bunch of APIs. And that's where Dynamic Workers shine.
https://t.co/i6Gz8xz9BZ
[0] https://t.co/IiGLvkrHnA
[1] https://t.co/cuZkcinSFk
[2] https://t.co/DAlLnPKdW2
We’re introducing Dynamic Workers, which allow you to execute AI-generated code in secure, lightweight isolates. This approach is 100 times faster than traditional containers. https://t.co/c36Vkb7I0R
To show this off I slopforked the @opencode server so it runs entirely inside a @CloudflareDev Durable Object.
No servers. No containers. Hibernates when idle. Try it:
opencode attach https://t.co/z0pWVur718
Code: https://t.co/AP1rx5wQ3y
Soudain je me souviens du traitement en France de la #copyrightdirective en 2019. Idem : la presse était pour, par intérêt direct parce que les ayants-droit lui avaient donné un bout de lard. Résultat, 100 % de la couverture en faveur. 1/2
Résumé de 100 % de la couverture média généralistes de la loi sur l'interdiction des réseaux sociaux aux moins de 15 ans : "super". Unanimité totale, y compris dans les émissions de débats, 0 recul, 0 contradiction. Les seules critiques regrettent que ça n'aille pas assez loin.
A good plan with phases allows one to limit the context to 40/50% per phase most of the time. That is where in my experience all models perform their best.
Someone woke up yesterday, decided plan mode wasnt important, and then a bunch of people on the internet decided to broadcast it in the echo chamber.
Stop trusting these people.
They are not experts.
I'm back to coding full time for work! I have re-joined @Cloudflare after I left there 9 years ago. I'll be working on their container product. Couldn't be happier to be working on problems of that scale again.
Some early HC employees will probably remember me joking that it was my divine mission to eliminate YAML from the world. I joked I started HC only to kill YAML. Like, back in 2013. And we (as an industry) were so close! Then Kubernetes came out and fucked it all up.
@jonas@MikeNomitch_CF@tqbf I'm not in Vegas but I would love to learn more about what you are building and how we can help. I'll DM you and we can find some time.