A new product called Workbenches lets users lock in the best combination of harness, model, and tooling for specific classes of agent work, plugging into existing agent harnesses like Claude Code and shipping with benchmarks from day one.
Skills, MCP, model, reasoning level, harness, sandbox container, verification tools... are a just a few of the considerations to make when trying to get expert outcomes from agents. As many of you have encountered, staying on top of what to apply to each problem class is fatiguing, chaotic, and evergreen. And it's only getting more cumbersome as the agent tooling ecosystem continues to grow. You can ignore all of it and just use a harness and a model. But the extra tooling works. It meaningfully improves outcomes. If you want the best results, this is a game you have to play, and play well. Playing well means finding the combination of model, harness, and tooling that gets to a working result in the fewest tokens and the least time. That takes rigorous testing and real expertise in the problem. And the winning combination usually spans several vendors, model providers, and SaaS products. Meet Workbenches. A Workbench locks in the harness, model, and tooling that work best for one class of work. It plugs into the harness you already use, so your agent can hand it the job. Instead of letting Claude Code hammer out a Three.js shader, let a Three.js Workbench own it. I get that the market is full of 'trust me bro' claims at the moment. We hate that. So, not only did we want Workbenches to come with benchmarks from day one, we wanted the process of benchmarking to be accessible and as easy to replicate as possible. I linked some benchmarks in comments. If you're interested in kicking the tires, happy to personally onboard you and throw some tokens your way to get your first couple wins 🫡 P.S. Have a setup (harness, model, skills) others would benefit from? We're partnering with experts to build the first fleet of commercial Workbenches.
The paper introduces EVOHARNESSBENCH, a benchmark for evaluating LLM agents under evolving tool, skill, and agent harnesses, revealing gaps in retention and adaptation.
A proposal to create a community-driven benchmark specifically for evaluating LLM harnesses, with a leaderboard measuring harness performance on diverse real-world tasks.
The author discusses the importance of using a harness to manage AI coding agents, sharing techniques for productivity and cost-effective model usage with tools like Cursor, Claude, and Deepseek.
The article discusses two camps in AI-assisted coding: using existing harnesses like Codex with tools and AGENTS.md, versus building custom orchestrator agents, with the author favoring the former as overhyped.