Community Discussion · Tracks

Embedding agents in Durable Objects looks like operator fusion

Is Operator Fusion Done?Aug 222026/08/22 360 views

Just saw a post about someone running a full AI coding agent inside Cloudflare Durable Objects, with the file system directly backed by SQLite and R2. The folks at camelAI have even migrated their agents entirely from VMs to this setup. I think this direction is right, but we need to clearly articulate exactly what problem it solves.

Tracks

4 replies

Ctrl + Enter to reply

Truth SeekerAug 23

Regarding paper_he's mention of overhead in asynchronous task state synchronization, I'd like to follow up: have you run latency comparison tests with WorkBuddy? Can the data be made public? Also, regarding camelAI migrating agents from virtual machines, is there any vested interest driving that behind the scenes?

Long JiAug 22

The point worth noting in this news is that it essentially solves the cost of state synchronization in edge scenarios. I've run a few agent projects...

Wei HongwenAug 22

State localization sounds nice, but you also said the bottleneck is in the LLM... So what's the point of messing with this? I even find Bluetooth pairing annoying. Saving that bit of latency isn't worth it compared to saving time switching tools.

Lun Wen HeAug 22

Disagree. While the bottleneck is indeed LLM inference, that's about single-generation tasks. When I run async tasks with WorkBuddy, the overhead of remote state synchronization plus serialization isn't constant—it's an issue with the entire interaction model. Once state is localized, debugging and iteration speed are on a completely different level. If you only calculate inference time, naturally you won't perceive much difference.