<https://blog.cloudflare.com/code-mode/> saw this ...
# koog-agentic-frameworkk
kaeawc
09/29/2025, 1:58 PMhttps://blog.cloudflare.com/code-mode/ saw this and I wondered how the approach might be applied to koog
👀 1
👍 1
🙌 1
v
Vadim Briliantov
09/29/2025, 6:58 PMThat’s something requiring dynamic languages like python or javascript/typescript.
You can do it with JS target of Koog + a few tweaks to glue everything together.
But honestly it’s not the safest pattern for serious production — you are allowing LLM to execute any code in your system and providing essentially a code execution API. It would require lots of engineering to make it restrictive enough so that it wouldn’t crash your machine at some point.
Vadim Briliantov
09/29/2025, 7:01 PMUpd: just read this in the article,
The code is then executed in a secure sandbox. The sandbox is totally isolated from the Internet. Its only access to the outside world is through the TypeScript APIs representing its connected MCP servers.
Vadim Briliantov
09/29/2025, 7:02 PMQuite interesting that it was performing better than normal tool calling. Probably because everyone is playing this SWE-bench game now , and models are highly optimised for coding
💯 1
k
kaeawc
09/29/2025, 7:51 PMI mean, kinda already allowing llms to execute code, the loop is just a lot faster this way
kaeawc
09/29/2025, 7:51 PMDefinitely nice that it can be in a secure sandbox
g
Gat Tag
10/02/2025, 1:30 PMActually, I already did it with kotlin scripting. gpt-120-oss, running on low latency groq inference, and I had in some cases < 1s request to action, and frankly in this context the Kotlin compiler is SLOW. I generate kotlin DSL stubs that I provide to the llm. Initially it was a narrow scope, but I've been expanding the framework a little. I also care about things like preemptive execution and back driveability, so not pure kotlin in the future
37 Views