The harness moved the bill more than the model did

Swap only the orchestration layer, holding the six models, the 22 locked tasks, the judge panel, and the price table fixed, and the bill falls 41%. That is the result Writer’s team reports in The Harness Effect, a controlled study that freezes everything except the software wrapped around the model. Cost per task went $0.21 to $0.12, tokens per task 14.2k to 8.8k (down 38%), median wall-clock 48s to 27s, with task-completion quality flat at 0.78 to 0.81. On this workload the orchestration layer moved cost per task more than switching between the cheapest and the most expensive model on the menu did.

The mechanism is unglamorous, and worth stating in full because every term is code, not model. A naive agent loop replays the full transcript each turn, so input tokens grow quadratically in turn count; agent workloads run input-to-output ratios near 100:1, so that replay is nearly the whole bill. Writer’s harness holds a byte-stable prefix (tool schemas, stable system prompt, append-only transcript) so that 99.9% of prompt tokens on an identical-prefix call served from cache at roughly a tenth of list price. It then compacts old history into typed checkpoints instead of middle-truncating, offloads bulky tool output to files the model reads by pointer, and classifies every failure before deciding whether to retry. Six mechanism families, one lever: cache the stable, compact the old, offload the bulky, suspend the waiting, bound the failing.

Two findings sharpen it. Efficiency is model-invariant, with all six models getting 33 to 61% cheaper and no exceptions, while quality gains track baseline capability almost perfectly (r=0.99 across the six). Stronger models convert orchestration structure into quality; weaker ones get overwhelmed by it, and all seven quality regressions landed on the three smallest models in the orchestration-heaviest categories. The authors call this harness leverage, and it carries a blunt corollary: the one net-new capability, delegated sub-agents, cleared a usable reliability threshold (0.85 and up) only on the two strongest models. A harness is a contract the model has to be strong enough to honor.

If Writer prices the harness, From Prompts to Contracts audits it. Kim and Ahn start from the same premise, that behavior living in a prompt is behavior you cannot guarantee, and move it into code: manifests, schemas, and validators around a replaceable model-composition boundary. Holding the model fixed and varying only the enforcement layer, prompt instructions alone let recommendation-language and internal-trace leaks reach the reader, while the code-owned validators blocked them across all 270 composition-boundary runs on three hosted models. A bolt-on external guardrail blocked the violations too but over-refused, dropping utility to 88/120 where the code-owned harness held full utility at 120/120. The guarantees you can actually ship are the ones the harness owns, not the ones you ask the model for.

The same logic, structure beating a longer single trajectory, is showing up in retrieval. WebSwarm replaces the lone ReAct agent, one long context trying to be both deep and wide, with recursive delegation: each search node carries a local objective, either answers it or spawns children, and returns evidence upward for the parent to expand or aggregate. It beats single-agent and flat multi-agent baselines on BrowseComp-Plus, WideSearch, and DeepWideSearch. ProjAgent targets a different retrieval blind spot: lexical, structural, and semantic similarity all miss repository functions that implement the same procedural logic under different names. It decomposes a target function into reasoning steps, retrieves functions with similar procedural behavior at each step, and folds that into ordinary semantic retrieval plus a static-analysis repair loop, reaching 41.14% Pass@1 on REPOCOD, ahead of retrieval-based baselines. Procedural similarity is a retrieval axis nobody was indexing.

All of it assumes you can see what the harness did. Mindwalk, a Show HN posted this morning, replays a coding agent’s session on a 3D map of the codebase. It is early and unproven, four points and no comments when I pulled it, but pointed at the right gap: the trace is where the token bill, the contract violations, and the retrieval misses all become legible after the fact.

What to watch: Writer’s swap is one harness, one vendor, n=22, with quality directional at that sample size. The claim that orchestration outweighs model choice as a cost lever needs a cross-harness, cross-workload replication, ideally on the long-horizon coding benchmarks the paper brackets out. If it holds there, the procurement question stops being which model you rent and starts being whose harness you run.