Apple Research Finds Minimal-Harness Coding Agents Match Complex MLE Orchestrators

Recent autonomous machine learning engineering (MLE) agents often rely on elaborate harnesses, including multi-agent orchestrators and dedicated retrieval subagents, to overcome perceived limitations in long-horizon cycles. However, a new study from Apple and EPFL challenges the necessity of this complexity. The researchers found that when using the same frontier LLM backbone and equal time budgets, open-source state-of-the-art harnesses provide no advantages over a single session of a minimal-harness coding agent baseline.

Minimal Harnesses vs. Complex Orchestrators

The study argues that the LLM backbone is the primary driver of performance, rendering additional machinery layers redundant in the coding agent setting. By comparing complex systems against agents with direct access to execution environments via basic primitives like read, write, and bash, the authors demonstrate that hand-crafted harnesses yield poor returns for current MLE benchmarks. This suggests that effort spent elaborating these systems may be misplaced when strong models are already available.