Compressing local coding agents

Goal

Software teams everywhere now rent coding agents by the token, sending their code to someone else’s servers and paying per use. Agents running on the company’s own hardware would cut those bills, keep the code at home, and keep working when the internet does not. The open question is how much agent fits on a single workstation.

Report

A 24GB workstation is on order and mid-sized open models are the starting point, the class rumored to run well in that much memory. Floating the idea among practitioners drew immediate enthusiasm, a good sign this itch is widely shared.

There is a deeper reason to be optimistic. Local AI tooling is fragmented across vendors and SDKs in ways that exhaust humans, and agents are interesting precisely because they have the stamina to grind through that fragmentation where people burn out. The thread has already spun out a poster proposal to D3A: Compressing Local Coding Agents: Directions and Open Questions.