Boris Kwemo

Letting AI agents run without "going rogue".

Building an open-source sandbox for Claude Code, Codex and others.

Ex VC-backed Founder of ConvertMate

[Boris Kwemo @boriskwemo]

I put Claude Code in auto mode and let it work.

The one thing it could not do on its own: reach a host I had not approved. It asked, I said yes in the Mac app, and it finished the job.

That is Fenzi: your AI coding agent in a sandbox on your laptop. One real run, cut down:

[Boris Kwemo @boriskwemo]

@SeHozaifa Don't pick an AI model, have a process that doesn't rely on the model that much

[Boris Kwemo @boriskwemo]

@krishnanrohit Hmmm that is a very very good question. Probably the most important question. It's difficult to say because I think the biggest impact is on the back-end and tedious processes users don't see. And widely used software were already v good for the most part

[Boris Kwemo @boriskwemo]

@PovilasKorop Totally fair split. I'd add a third group: people who mix them. Cheap model for the boring stuff like renames and tests, the big one only for planning and the hairy bugs. Do you see many people doing that, or do they pick a side?

[Boris Kwemo @boriskwemo]

@hwchase17 The tricky bit is who decides a run went well. A passing test is easy. A run that passed but took a weird path is harder. Are people mostly labeling by hand, or letting evals pick the keepers?

[Boris Kwemo @boriskwemo]

@thorstenball Sure, when you have a strong structure and clearly defined processes, it makes code review easier and the whole CI/CD faster

[Boris Kwemo @boriskwemo]

@kentcdodds Tips: Add terms like "Extremely creative"

[Boris Kwemo @boriskwemo]

@oegerikus 100% agree. Having a shared policy and sandboxing for every dev is becoming more and more important

[Boris Kwemo @boriskwemo]

@rdominguezibar But why the hell would they give tools to AI to do these kind of actions?

[Boris Kwemo @boriskwemo]

@SeHozaifa @KhaliqHussainnn One model a week

[Boris Kwemo @boriskwemo]

@RahulbeGupta Google

[Boris Kwemo @boriskwemo]

@sqs Love that it's deterministic. Does the rule look at the arguments too?

[Boris Kwemo @boriskwemo]

This is absolutely genius

[dvb @notdvb]

are you fucking kidding me?

https://t.co/La3RH0f4f0

[Boris Kwemo @boriskwemo]

@notdvb This is a masterpiece LOL I laughed so much

[Boris Kwemo @boriskwemo]

@OpenAIDevs That was painful, 5 mins in and still only jokes :(

[Boris Kwemo @boriskwemo]

Honest poll for anyone using coding agents:

A) approve every command as it comes
B) sandbox it, let it run, review at the end

Which one are you, and what pushed you there?

[Boris Kwemo @boriskwemo]

@ZackKorman Could not agree more. 0% accountability, 100% story-telling.

[Boris Kwemo @boriskwemo]

@sardor Man to man lmao

[Boris Kwemo @boriskwemo]

@ZackKorman Are they going to go through the entire chain? It's fine to audit their contractors, but if the contractors' contractors use DeepSeek, will they really know?

[Boris Kwemo @boriskwemo]

@aditikbuilds Why is it a problem?