Harness, not model, now decides agent performance and cost

Harness design, not model choice, is now the variable separating working agents from broken ones. ARC-AGI jumped from roughly 30% with Claude Opus to 95% once teams changed scaffolding and context management. Treat the number cautiously—benchmark contamination came up—but the direction is consistent.

At YC, agents evolved from a one-size-fits-all system to QM: the agent brain lives in Postgres, sandboxes are disposable resources, and agents choose compute and model runtime. Payoff: operational control and shared context. Tradeoff: permissions bound what agents can know, and human review still gates database writes.

Prime Agent shows the next shift: self-improving harnesses that update prompts, memory, and subagent specs, and run multi-day research with hundreds of agents. The evidence is promising but anecdotal; a $5,000 Aider run and one 'cheating' first result should temper enthusiasm. Long-horizon evaluation needs fixed budgets, not hand-picked winners.

Open Jarvis pushes local inference as a privacy and cost escape hatch, claiming local models are 6 to 12 months behind frontier and 800x cheaper to run. That is more speculative. Leaders should track it, but not bet core architecture on local-only stacks yet. Watch how quickly self-improving harnesses and permission systems become differentiators.