Agent performance depends on the harness as much as the model. JIT-Agent is a trained meta-agent that takes a task description and emits a task-specific harness wrapping any off-the-shelf agentic LLM. It can repair harnesses and evolve them from an archive of past configurations.\n\nThe authors report that DeepSeek-V4-Flash with JIT-Agent surpasses GPT-5.6 by 9.1 points on DeepSearchQA and 4.3 on OdysseyBench, while GLM-5.2 gains up to 20.2 points. JIT-Agent-generated harnesses are competitive with OpenCode and Claude Code across controlled evaluations and improve DeepSeek V4, Mimo-V2.5, and Qwen3.6 model families.\n\nThe interesting claim is that harness construction is a trainable, transferable skill, orthogonal to model scaling, rather than a manual per-task engineering effort.
Generating Agent Harnesses on Demand Improves Models Across Benchmarks
JIT-Agent dynamically builds and repairs task-specific agent scaffolding, producing reported gains across several model families and agent benchmarks.
Top University
Guibin Zhang · Leo Lu · Fangzhou Xie · Kang Zhu · Junhao Wang · Zhifei Xie · +10 more
LV-NUS Lab
Research Digest··2 min read
Thread:Agent Harness Optimization
In one line
JIT-Agent trains a meta-model to generate task-specific agent harnesses on the fly, improving off-the-shelf LLMs' agent performance.
Before you read this
Three ideas this paper builds onagent harnessA system that structures, verifies, monitors, or controls how an AI agent interacts with external tools, environments, or execution runs.150 papers →harness evolutionIteratively adapting an agent's environment, prompts, or evaluation setup to improve task performance and generalization.56 papers →deep researchA multi-step AI system that autonomously gathers, evaluates, and synthesizes evidence from diverse sources to produce comprehensive reports.21 papers →
Read it as
The authors train a model to generate an agent’s harness—the memory, planning, action, and tool-orchestration machinery surrounding an underlying language model—for each task at execution time.
The authors represent an as a machine-generatable artifact governed by a fixed four-module protocol.
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.