The authors built Meta-Harness, an automated system in which Claude Code proposes and revises domain-specific Python harnesses using previous source code, execution traces, and scoring artifacts.
Joint harness search balances model accuracy, safety, and token use
Across 17 domains and 12 models, jointly optimizing the code around language models outperformed sequential and accuracy-only approaches.
Big Tech
Subhojyoti Mukherjee · Md Mehrab Tanjim
Adobe Research
Research Digest··2 min read
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.176 papers →harness evolutionIteratively adapting an agent's environment, prompts, or evaluation setup to improve task performance and generalization.66 papers →safety alignmentProcess of training models to recognize and refuse harmful requests, ensuring outputs remain safe.106 papers →
Mukherjee and Tanjim treat an LLM harness, the code that builds prompts, routes calls, and parses responses, as a multi-objective optimization problem.
How it works
before and afterThe Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.