Why controlling what AI agents compute is architecturally hard: GPUs and harnesses live apart
jd_pressman · x · 2026-10-24
In a discussion with @lumpenspace and @dcolascione, jdpressman argues that any scheme requiring meaningful control over GPU computation or execution permission on GPU boxes is impractical given how compute is actually organized.
He explains that inference GPUs are typically separated from the machine running the agent harness, so an "AI agent" may physically span multiple GPUs in different locations, none colocated with the CPU executing its commands — making fine-grained control of agent compute architecturally difficult.
Related event: Debate: Can Air-Gapped GPUs Leak Data via Thermal Side Channels?(3 posts)→
More from coding & agent
- Builder pairs Grok Bot for firmware with locally-run Hermes Agent inside a Stack-chan robot — alexcovo_eth · 2026-10-24
- 30% of Git modelled in Lean 4 with coding agents, 8,000+ theorems proven, 100+ potential bugs found — kfountou · 2026-10-24
- REA-Prompts hits 1,700+ GitHub stars in days: prompts that make coding agents reverse-engineer any software — Promptmethus · 2026-10-24
- Kimi K3.1 Pro stumbles on a real agent coding task: ordering errors sink the build — QuixiAI · 2026-10-24
- Veteran's advice: stop relying on specs alone, just talk to your LLM — MoonL88537 · 2026-10-24
- Agent orchestration tip: skip goal, use a manager agent with scheduled polling — pvncher · 2026-10-24