Poolside CTO Eiso Kant proposes replacing LLM tool calling with micro-VMs

3:58 PM · Jul 20, 2026

I would love a company with LLM training chops (ahem ahem) to put this to the test From first principles optimal seems something like this - input tokens are just “events” - things that happened in the world - output tokens are code - code runs in a repl so you get to reference variables from previous turns - the output language supports streaming data types and is parsed on the fly - so you can even stream token-by-token to (in our case) itx.slack.getThreadStream() or sth It would pin inference to a repl session but with websockets inference APIs, realtime voice models etc there increasingly is some pinning of non-gpu workloads around the inference anyway

ON Orion Night@orionintx

4:04 PM · Jul 20, 2026

@eisokant tool calling is just function pointers with extra serialization and a per-token bill

EK Eiso Kant@eisokant

3:41 PM · Jul 20, 2026

1991TECH

Been saying a version of this for 2 years now. We should give each model its own micro VM with its own code base and just stop doing tool calling all together. Models are more effective and efficient with code.

FI finbarr@finbarrtimbers

4:29 PM · Jul 20, 2026

656TECH

@eisokant Yeah imo this is the way to go. Tool calling is a hack.

Poolside CTO Eiso Kant proposes replacing LLM tool calling with micro-VMs

Combined views

Poolside CTO Eiso Kant proposes replacing LLM tool calling with micro-VMs

Reactions from ranked influencers

Combined views

Sentiment

Positive Read

Digg Deeper

Reactions from the X community

Sentiment

Positive Read

Digg Deeper