Beyond Model Choice: Harness-Mediated Capability Substitution in Agentic Software Engineering

Whether an open-weight language model can replace a proprietary frontier model for software engineering is not a question that can be settled from model rankings alone. Agentic coding performance depends on a coupled system: the model, contextual retrieval, reusable procedural skills, permitted tools, execution feedback, repair policy, and human approval. This paper develops a reproducible framework for evaluating harness-mediated capability substitution: the conditional use of lower-cost open-weight models when a fixed, verifiable engineering workflow can deliver output that meets an organization’s acceptance standard. The proposed architecture preserves a developer’s existing OpenCode harness while adding versioned skill contracts, risk-aware task routing, repository-grounded context packing, sandboxed execution, deterministic and visual verification, and telemetry for escalation and cost allocation. It introduces a historical merged-pull-request replay protocol with paired model comparisons, blinded reviewer judgments, hidden acceptance tests, design-system-specific oracles, and predeclared non-inferiority criteria. A cache-adjusted cost-per-success model incorporates token classes, retries, tools, infrastructure, reviewer time, maintenance, and failure externalities. An explicitly hypothetical calculation demonstrates why low inference prices do not necessarily imply low total engineering cost. The contribution is an implementable research design, analytical model, and deployment reference architecture; it does not report new benchmark runs or establish general performance parity. The central claim to be tested is narrower: for some structured, bounded, and well-verified classes of coding work, an open-weight model plus a capable harness may satisfy the same practical quality threshold as a frontier-model workflow at a lower risk-adjusted total cost.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-16
DOI
https://doi.org/10.5281/zenodo.24287644
Primary Topic
Software Engineering Research
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Beyond Model Choice: Harness-Mediated Capability Substitution in Agentic Software Engineering

Karan Chandra Dey
2 links
Zenodo (CERN European Organization for Nuclear Research)
Software Engineering Research
preprint

Beyond Model Choice: Harness-Mediated Capability Substitution in Agentic Software Engineering

Karan Chandra Dey
preprint en

Abstract

Whether an open-weight language model can replace a proprietary frontier model for software engineering is not a question that can be settled from model rankings alone. Agentic coding performance depends on a coupled system: the model, contextual retrieval, reusable procedural skills, permitted tools, execution feedback, repair policy, and human approval. This paper develops a reproducible framework for evaluating harness-mediated capability substitution: the conditional use of lower-cost open-weight models when a fixed, verifiable engineering workflow can deliver output that meets an organization’s acceptance standard. The proposed architecture preserves a developer’s existing OpenCode harness while adding versioned skill contracts, risk-aware task routing, repository-grounded context packing, sandboxed execution, deterministic and visual verification, and telemetry for escalation and cost allocation. It introduces a historical merged-pull-request replay protocol with paired model comparisons, blinded reviewer judgments, hidden acceptance tests, design-system-specific oracles, and predeclared non-inferiority criteria. A cache-adjusted cost-per-success model incorporates token classes, retries, tools, infrastructure, reviewer time, maintenance, and failure externalities. An explicitly hypothetical calculation demonstrates why low inference prices do not necessarily imply low total engineering cost. The contribution is an implementable research design, analytical model, and deployment reference architecture; it does not report new benchmark runs or establish general performance parity. The central claim to be tested is narrower: for some structured, bounded, and well-verified classes of coding work, an open-weight model plus a capable harness may satisfy the same practical quality threshold as a frontier-model workflow at a lower risk-adjusted total cost.

Zenodo (CERN European Organization for Nuclear Research)
Software Engineering Research
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon