Two prompts can request the same code change and produce the same correct patch, yet cause a coding agent to perform radically different kinds and amounts of work. We study this effect in a preregistered benchmark spanning 4,644 valid runs, 24 deterministic coding tasks, seven reasoning models, and two real agent harnesses.
The study investigates how individual components of a coding harness—planning, action space, and context management—affect autonomous coding agents’ performance. By fixing the execution loop and varying these components across 176 settings on SWE‑Bench Verified and Terminal‑Bench 2.1, the authors find that context management is most valuable when context windows are tight, staging rule‑based elision before LLM summarization yields the best efficiency, planning serves as an accuracy scaffold for weaker models and a cost saver for stronger ones, and predefined tools help models with limited bash skills while bash‑capable models benefit from a bash‑only interface. Trajectory‑level analysis shows that context management lengthens execution paths, planning alters where trajectories terminate, and the action space determines code granularity, offering a modular framework for future harness design.
arXiv:2603.19896v2 Announce Type: replace Abstract: Tool-using large language model (LLM) agents often face a fundamental tension between answer quality and execution cost. Fixed workflows are stable...
arXiv:2609.01437v1 Announce Type: cross Abstract: As agents move from research prototypes to deployed tools, their capability increasingly depends on model-external execution infrastructure, commonly...
The paper introduces Growing Harness, a training method that transforms recurring control logic in large language model agents into reusable executable code, reducing reliance on the model for task-specific decisions. By using strategy-free scaffolds, failure-guided code repair, and success-first gating, the approach learns a shared harness that improves performance across multiple benchmarks and model sizes. Experiments on BrowseComp-Plus and WebArena-Verified show significant gains in success rates and substantial reductions in LLM calls and inference cost compared to traditional tool‑calling agents.
The paper investigates cost-inefficient behaviors in coding agents, analyzing 1,200 trajectories from Claude Code and Mini-SWE-Agent on SWE-bench Verified. It identifies three main inefficiencies—subsumed retrieval, similar script generation, and test re-execution—that affect 79–98% of tasks and contribute up to 22.75% of costs. The study evaluates mitigation strategies, finding that developer-designed skills reduce costs by up to 41.73%, outperforming other approaches.
arXiv:2607. 06413v1 Announce Type: cross Abstract: Large language model coding agents increasingly perform open-ended data modeling and analysis.
The paper argues that evaluating agents as fixed models is insufficient, proposing instead to treat them as configurable systems. Using a new benchmark of four scientific tasks, the authors analyze how five configuration aspects—task information, reasoning, self‑verification, time budget, and backbone model—affect performance, noting that about 54% of outcome variance arises from run‑to‑run differences even with the same settings. The study finds that providing more task information has the strongest impact, while interactions among settings (e.g., extra time only helps with adequate information or model capability) and the choice of verification tools significantly shape agent behavior.
Agentic AI development today runs on token maxing: buying capability with tokens -- longer reasoning traces, more turns, wider tool payloads, bigger replayed contexts -- so tokens per task grow faster than task value. Falling per-token prices mask the pattern; total spend rises anyway.
The paper introduces Harness Primitives—reusable agent harness mechanisms mined from failed task trajectories—and a framework called STITCH that selects and composes these primitives into task‑specific harnesses at test time. This approach avoids generating or debugging harness code for each task, achieving up to 12‑point gains in task success over fixed harness baselines and outperforming human‑designed harnesses like Codex CLI. STITCH also demonstrates minimal test‑time overhead (2.7%) and scales efficiently with the size of the primitive library.
arXiv:2601. 14192v2 Announce Type: replace Abstract: Recent years have witnessed increasing interest in extending large language models into agentic systems.
arXiv:2511. 17006v2 Announce Type: replace Abstract: Scaling test-time computation has been extended from language model reasoning to tool-augmented agents, where scaling involves not only thinking in tokens but also acting via tool calls that directly constrain environmental interaction.