Editorial illustration for Better Harness updates add usage examples, chaining guide, and tool clarifications
Better-Harness Updates Boost AI Research Workflow
Better Harness updates add usage examples, chaining guide, and tool clarifications
The latest round of Better Harness updates is not a tweak. It’s a recalibration. We’ve added usage examples, a chaining guide, and clarified the tool suite, those edits crack open what was once opaque.
Why? Because running a harness blind leaves too much to guesswork. We tested the updated loop with Claude Sonnet 4.6 and Z.ai’s GLM-5 on a subset of evals.
Results confirmed what we suspected: clear examples and disambiguated tools make hill-climbing faster, not just cleaner. This is groundwork. A broader eval suite is already in motion across deepagents, aiming for public model profiles that capture each model’s idiosyncrasies.
The goal: turn nuance into a repeatable artifact. Read on for what changed, what held, and how the new harness sharpens the climb.
These updates strip away the friction. Examples now show you the exact shape of a working call. The chaining guide maps the flow from one turn to the next.
And the tool descriptions? Cleaner, sharper, no more guessing which harness does what. We put the revised loop to work with Claude Sonnet 4.6 and GLM‑5.
The results confirmed what we suspected: clarity breeds consistency. But this is just a snapshot. A broader effort is already running, generalizing Better-Harness across more models in deepagents, using a larger eval suite.
The real prize is a series of public model profiles, each tuned to our evals and capturing what makes that model tick. That is the point. Not a black-box leaderboard, but a nuanced map of behavior.
The harness is the tool. The profiles are the insight. What you build with them is up to you.
Related Reading
Research & Benchmarks Google's FACTS benchmark shows 70% factuality ceiling across four tests
Research & Benchmarks Databricks finds multi-step agents beat single-turn RAG by 21% to 38% on STaRK
Research & Benchmarks Nvidia's DLSS 4.5 beta adds 6x Multi Frame Generation for RTX 50 GPUs
From Archives Study finds ‘bot’ term used 16,232 times in 2.8M Telegram messages
From Archives Google AI Overviews answers 91% of test questions correctly after Gemini 3 update
Common Questions Answered
What specific updates were made to the Better-Harness tool suite?
The updates include concrete usage examples, a comprehensive step-by-step chaining guide, and a refreshed description of each component. These changes aim to reduce confusion and provide clearer guidance for researchers on how to effectively use and integrate the tool.
How did the team validate the effectiveness of the Better-Harness updates?
The team tested the updated approach with Claude Sonnet 4.6 and Z.ai's GLM-5 on a subset of their evaluations. They are also working on generalizing Better-Harness across multiple models using a larger evaluation suite to capture nuanced performance characteristics.
What is the core principle behind the Better-Harness approach?
The Better-Harness approach treats evaluations as training data for AI agents, similar to gradient-driven loops in classical machine learning. Each evaluation case provides a signal about whether the agent took the correct action, which can be fed back into the harness engineering process to improve performance.
Further Reading
- How to Build Tool Chaining — OneUptime
- Writing Effective Tools for Agents — Tool naming, schema design, error messages, and return value conventions — GitHub (ai-boost)
- Output variables with chained pipeline — Harness Developer Hub
- Realtime Eval Guide — OpenAI Developers
