Duo Lu, Andrew Crotty, Uğur Çetintemel
Reusable declarative primitives extracted from searched LLM harnesses can be catalogued and recompiled into working harnesses for unseen tasks at roughly one order of magnitude lower cost than running agentic harness search from scratch.
On four self-constructed workloads (finance retrieval, legal retrieval, math, polyreason) built from public benchmarks with 100-query search and in-distribution splits and 300-query cross-distribution splits, the authors run Meta-Harness search (10 iterations) per domain, extract typed primitives (skeleton, beliefs, policies, prompt templates) via a Claude Opus abstractor, then test (1) round-trip fidelity on 3 seeds x 3 compile replicates per domain (n=9 each) on 100-question held-out splits, and (2) four cross-domain primitive-reuse transfers (finance<->legal, math<->polyreason), comparing accuracy (gpt-5-mini judge) and per-question cost against brief-only and fully searched harnesses.
Round-trip fidelity is tight: accuracy changes at most +/-0.011 and cost within +/-0.013 cents/question. One recover-and-compile costs $2.3-2.8 and minutes versus $43-72 and 2.8-9.2 hours of wall time for full search. Reuse: legal->finance lifts accuracy 15.7 to 28.3 (searched reference 27.7); finance->legal lifts 43.4 to 66.3, recovering 76% of the gap to the searched 73.4. Transfer breaks for reasoning: math primitives applied to polyreason reach 34.9-38.7, below the brief-only 43.3, showing reuse depends on matched operating points. Scope conditions: four domains, one searcher, one frontier abstractor model, all hand-selected workloads, preliminary results only.
The authors label results preliminary; accuracy on finance (28-29%) and legal (67-73%) leaves large headroom, and the finance reuse result falls within run-to-run variation. The compiler draws on frontier model knowledge, which the authors acknowledge could confound fidelity claims (they partially control via cost matching). Only Meta-Harness is tested as the source of harnesses; transfer is tested only within task families (retrieval to retrieval, reasoning to reasoning); no formal equivalence rules exist for reordering steps; the reasoning-transfer failure shows the approach is not general. Unstated: the 100-query search and evaluation splits are tiny, judge-based accuracy (gpt-5-mini) introduces grading noise, and no code or catalog is described as released beyond a GitHub link to the searcher.
A skeptic would note the headline 'order of magnitude savings' compares full search against one cheap compile call after the search already happened, so the saving only materializes when reuse actually works, which it did not for reasoning transfer. Four domains, three seeds, and 100-300 question splits give low statistical power, and reported differences of a few accuracy points may be within seed noise. The workloads are bespoke constructions from benchmark bundles rather than real enterprise tasks, and the abstractor/compiler run on a frontier model whose priors could quietly rebuild good harnesses regardless of the catalog. This is a vision paper for a database-community research agenda with preliminary evidence attached.
Little or nothing yet for the AI economy in the next 12 to 24 months. If the approach matures, it would reduce the marginal cost of deploying agentic systems across related tasks by amortizing harness-search spend, but this is an early systems-research result on toy workloads, not a production cost curve.
Technical Capability: Harness engineering cost, not just model quality, is shown to be a real expense of agentic systems, and this paper argues that cost can be amortized via reusable declarative structure.
Labor & Knowledge Work: If harness search becomes cheap and catalogued, it shifts skill demand from prompt/harness tinkering toward catalog curation and compilation, a small but concrete reallocation of AI engineering work.