QuarryIf the model stays the same, does the harness around it change what an agent costs?
From tokenmaxxing to valuemaxxing: the hidden economics of the AI harness
Same model, same 29 tasks, seven agent harnesses. Completion varied by three tasks; estimated cost per successful task varied 2.5 times. The harness is becoming a material variable in the economics of agents.
Jaidev Karthickeyan5 min readData: Composio Bench; Liu and Han (2026)
What happened
Take one model, give it the same 29 tasks, and change nothing but the software wrapped around it. The bill moves two and a half times.
That is roughly what Composio found when it ran GPT-6 Astra through a line-up of agent harnesses. Completion barely moved, from 69.0% to 79.3%: Cline finished 23 tasks; Codex, Hermes Agent and Command Code 22; Claude Code 21; OpenCode and Pi Agent 20. Cost moved a lot. Composio estimated cost per successful task at $1.06 with Pi Agent and $2.62 with Claude Code, which also finished behind four other harnesses. When harnesses failed, they burned three to five times as many tokens getting there.
Updated 3 November 2026: Composio has since re-scored the run, which lifted each harness by one task, and added Cline. The completion figures here are the current ones; the cost estimates are from before the re-scoring.
Same brain, same exam, different bill. My claim is that the difference is the harness.
Going to ground
A harness is everything between the model and the work: the loop that decides what to try next, the context it carries, the tools it can reach, the checks it runs and the point at which it declares victory. The model supplies the thinking. The harness decides how that thinking gets spent.
Every one of those six jobs costs tokens, and every one can be done well or badly. That is where the bill comes from.
First order
Tokenmaxxing is not valuemaxxing. The first generation of AI economics was a game of tokens: price per million, tokens per second, cache-hit rates. They matter, the way fuel economy matters. But nobody buys a lorry for its fuel economy. They buy it to deliver things.
The benchmark makes the point neatly. The cheapest harness per success, Pi Agent, finished joint last on completion. The most expensive, Claude Code, didn't finish first either. A system can use fewer tokens and fail more often, or make fewer calls and need more human rescue.
The question that matters is not how many tokens an agent eats. It is what it costs to get a job done, checked and right.
Second order
The harness sets the price of execution. Five mechanisms do the work.
Context. Every call reprocesses whatever is in the window. Resend the whole history and every tool schema and you pay for it again and again. Trim too hard and the model forgets what it needed. The skill is just enough context, not the least.
Orchestration. Bad orchestration duplicates work and hands it around like a parcel nobody signed for. Good orchestration routes each step to the right model, reuses results and runs independent steps at once. More agents is not better; every extra one adds a meeting.
State. Long workflows have to remember where they are. Without checkpoints, an agent rebuilds old decisions and redoes finished work, and you pay twice.
Verification. An agent saying it's done is not proof that it is. Check too little and errors slip through; check too much and overhead eats the savings. The right amount depends on what a failure costs.
People. Human review is part of the machine, not a cost outside it. The measure that matters is not the share of tasks labelled autonomous. It is how much human effort a reliable result still needs.
Third order
The fight moves to the execution layer. Model makers compete on capability, speed and price. But a capable model is not the same as an affordable deployment. The harness sits next to the work, where the model meets the company's data, tools and permissions, and that is a good place to stand.
It offers three advantages. Portability: a well-built harness lets you swap models as prices and capabilities change. Learning: its telemetry shows which tasks fail, which tools cause friction and where context is wasted. Control: whoever runs the execution layer shapes how intelligence is used and which systems it touches.
That doesn't make the harness a moat. Much of it is copyable, and model providers are building their own. The defensible part is the combination: integration with the workflow, operating data, reliability and governance, accumulated over time.
Ears up: what to watch
Cost per completed task, not cost per model call.
Tokens, tool calls and model turns per verified outcome.
Failure and recovery rates, and human intervention per workflow.
Latency, at the median and the 95th percentile.
Whether rankings survive a change of model. So far they don't: Codex tied for first on GPT-6 Astra but came last on success with Kimi K3, and on DeepSeek V4 Flash a different harness won on each of success, cost and speed. There may be no best harness, only the best fit.
Bone of contention
The sharpest counterpoint to my argument comes from a preprint with an excellent title: What Does a Harness Buy? Tokens, Mostly, by Liu and Han. On a 447-task pool, the heaviest harness, Claude Code, and the lightest, mini-SWE-agent, finished within five points of each other on both models tested; OpenCode trailed by up to nine, partly because runs hit an output cap. On a 45-task run, cost per trial differed by about three times.
The authors blame much of the gap on resending system prompts and tool schemas at every step: about 16,600 tokens of fixed overhead per step for Claude Code against about 800 for mini-SWE-agent. That explanation is disputed, and their own numbers show it isn't the whole story, because OpenCode carried more overhead than mini-SWE-agent and was still cheaper per trial. They also ran only two locally hosted Qwen models on the large pool.
Composio's benchmark has limits too: 29 tasks, so one task moves a harness 3.4 points; costs estimated from list prices; and costs published for only two harnesses, before the re-scoring. So harness engineering may buy big savings without buying better results. A pricier harness isn't automatically smarter, and a cleverer one isn't automatically more productive. The test is whether the extra complexity buys reliability, efficiency or output you can measure.
The model decides how much intelligence is available. The harness decides how it gets spent. Tokenmaxxing counts the fuel; valuemaxxing counts the deliveries.