Coding agents are getting cheaper the boring way: routing the work, not rebuilding the harness
A scouting agent matches the best single model at about a fifth of the cost, and a timeout cascade cuts GPU costs by a third. The harness itself buys far less than its price suggests.
- In a preprint, a scouting system called SuperScout matches the best single model's solve rate on SWE-bench Pro (159 of 266) at about a fifth of the total cost per solve. Self-reported m:…0961
- In a preprint, a timeout-based cascade cuts GPU-hour costs by 33% on SWE-bench Verified without sacrificing accuracy. Self-reported m:…0960
- In a preprint, harness swaps change task outcomes at a rate comparable to simple reruns, and inference cost differences between harnesses vary by model, ranging from under 10% to 3-4x. Self-reported m:…8345
- In a preprint, SWE-Growth improves Qwen3-Coder-30B-A3B-Instruct performance on SWE-bench Pro from 27.7% to 31.7%. Self-reported m:…4031
What changed
The cheapest coding agent in this week's preprints is not a new model. It is a routing decision. In a preprint, a system called SuperScout matches the best single model's solve rate on SWE-bench Pro, at 159 of 266, at about a fifth of the total cost per solve. The trick is a searcher that scouts the repository first and hands a structured plan to the model that does the work, so the expensive model does not have to discover the codebase itself. m:…0961
A second preprint reaches a similar place from the other direction. Small language models achieve high accuracy on easy and medium complexity coding tasks, with the performance gap to large models concentrated in hard tasks. A timeout-based cascade strategy exploits that split: it starts cheap, and only escalates when the clock says the task is hard. The result is a 33% reduction in GPU-hour costs on SWE-bench Verified without sacrificing accuracy, and simple timeout cascades perform comparably to oracle-style task routers for coding agents. m:…0960
The harness question
That puts a third preprint in an awkward spot. It asks what a harness actually buys, and the answer is mostly tokens. Harness swaps change task outcomes at a rate comparable to simple reruns, which suggests that much of what looks like a harness improvement is ordinary variance. Harness performance is model-dependent, with specific pairings showing significant performance swings, so a harness that helps one model can hurt another. And inference cost differences between harnesses vary by model, ranging from under 10% to 3-4x, which means the same harness can be cheap on one model and expensive on another. m:…8345
The implication is that teams shopping for a harness are asking the wrong question first. The evidence here says the money is in deciding which model sees which task, and how much scouting happens before the expensive model starts. A harness that does not improve that decision is mostly charging for the privilege of running the same model with different scaffolding. That does not make harnesses useless, but it does make their claims harder to take at face value without a rerun baseline. m:…8345
Training still matters
Routing is not the only lever. A fourth preprint, on a method called SWE-Growth, improves Qwen3-Coder-30B-A3B-Instruct performance on SWE-bench Pro from 27.7% to 31.7%, and improves performance on TerminalBench 2.1 from 11.0% to 17.2%. The method uses teacher-verified outcomes to guide self-play training, and it exceeds student-only self-play by 2.6 and 4.1 percentage points on SWE-bench Pro and TerminalBench 2.1, respectively. In other words, a smaller model can be taught to do more of the work itself, which is a different route to the same destination: fewer expensive calls. m:…4031
There is also a quality thread running alongside the cost thread. A preprint on metrics-in-the-loop finds that its intervention significantly reduces code erosion in autonomous coding agents, and that complexity feedback does not negatively impact held-out correctness of agent-written code. Timely, actionable measurements can redirect agents before weaknesses become foundational. That matters because a cheap agent that writes tangled code is not cheap for long; the repair bill arrives later, in a form no routing table captures. m:…9774
What to do with this
For a team running coding agents at any real volume, the practical read is straightforward. First, measure the cost per solve, not the cost per call, because SuperScout's advantage only shows up at that level. Second, test whether a scouting step or a timeout cascade beats your current setup on your own tasks, since the harness preprint shows results swing by model pairing. Third, keep a rerun baseline next to every harness comparison, because without one you cannot tell a real improvement from noise. None of this requires a new model. It requires treating the agent's workflow as the product, and the model as one expensive ingredient inside it. m:…0961 m:…8345
What is still unresolved
- Whether SuperScout's cost advantage holds on private codebases rather than SWE-bench Pro tasks.
- Whether timeout cascades keep their accuracy advantage when tasks are not cleanly split into easy, medium and hard.
- How much of the harness variance finding generalises beyond the models and tasks tested in the preprint.
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.