Harness²
Harness² is a plug-and-play framework designed by Google Research that enables large language model agents to recursively refine and optimize their execution harnesses—including prompts, tools, memory, and skills—across tasks encountered in an open world. It uses a dual recursion approach of task-level refinement and harnessing-level learning without requiring ground-truth verification, improving agent performance on professional benchmarks.
Ratings & Reviews
Write a reviewBe the first to share a thoughtful review.
Information
Around the Web
Rewriting its own harness between runs without a grader telling it what correct looks like is the part I find hard to believe, so I want to see whether the accumulated experience actually transfers between unrelated tasks.
karimbenkeroum · Imported from Product Hunt
I saw the runners are portable, whereas the judge goes through vertex. Is bring your own task inference a planned feature ? Will the open PRs for other vendor support eventually be accepted ?
amartya_gaur1 · Imported from Product Hunt
Agents improving their own setup is very cool . How do you prevent them from making changes that can make things worse?
rania_rimali · Imported from Product Hunt
Love that this is open source. My one worry with self evolving harnesses is the loop learning the wrong lesson. Without a ground truth verifier, what stops one lucky run from becoming a permanent rule? We run new rules in shadow mode first and they only start blocking after they prove they help on a few tasks. curious if you do somthing similar
ali_shaheen · Imported from Product Hunt