A harness rewrite cut agent token use nearly in half

SoL-Pi scales automated research loops over harness design instead of over the underlying model. Tested on the 51-task EdgeBench evaluation with GPT-5.6 Sol and Opus 5, it matched Pi's task performance while cutting recorded token traffic by 44.7-49.0% and API cost by about a third, per the SoL-Pi paper.

That's the part worth sitting with if you run agents continuously rather than in one-off sessions. The authors put the savings at $8.75 to $13.50 per hour against native Codex and Claude Code harnesses, tested at equal task performance rather than a degraded one.

Four mechanisms survived the selection process: action execution, context compaction, observation handling, and delegated reading. None of them touch the model weights. They're about what the harness does with tokens before and after the model sees them, how it trims context, how it hands off reading tasks instead of stuffing everything into one call.

The paper doesn't say whether these four mechanisms transfer cleanly to harnesses built for different task shapes than EdgeBench's 51 tasks, or whether the savings compound or plateau at larger scale. It also doesn't say how much engineering effort building a harness like this takes compared to just switching models.

Still, the framing is useful even without full generalization: EdgeBench is a fixed benchmark, and the real test is whether teams running agents in production see the same 44.7-49.0% drop once their task mix stops looking like a research paper's.