AI Pulse · Under the Radar

Most teams picking an AI agent worry about which model to use. Turns out the harness, the scaffolding that wraps the model and handles tool calls, matters just as much for cost.

Composio ran the same 28 tasks through three different harnesses, all using the same Kimi K3 model. Success rates stayed roughly even. Token consumption did not. The median task burned 61,000 tokens when run through Kimi Code’s harness, 138,000 through Hermes, and 340,000 through Claude Code’s harness. That’s a 5.6x spread on identical work with identical outcomes. The numbers come from Composio’s own benchmark thread and have not been independently reproduced yet, but the finding is drawing attention from builders this week because the gap is large and the implication is straightforward: two agents can deliver the same result at wildly different cost depending on how they are wired.

Why this matters if you run agents in production

If you are running customer support bots, data extraction pipelines, or any workflow where an agent fires dozens or hundreds of times a day, a 3x or 6x token difference compounds fast. A harness that loops unnecessarily, re-reads context, or structures tool calls poorly will chew through your budget even if the underlying model is cheap. This is exactly the kind of efficiency question we bake into systems at the Omni Command Centre, where cost per task and harness behaviour sit alongside model choice in the design process. The takeaway is not to avoid any particular harness. It is to measure token use in your own workflows, compare options, and treat the scaffolding as a variable you control, not a given.