Many LLM-based optimization algorithms have been proposed recently....
Shangyin Tan@ShangyinT
261 views
Jul 23, 2026
~2 min read
Advertisement
1
Many LLM-based optimization algorithms have been proposed recently. We find that GEPA, AutoResearch, and Meta-Harness each win on different tasks and no single LLM optimizer wins everywhere.
How do we leverage the complementary performance profiles of these optimization algorithms?
Today, we release optimize_anything omni: a meta-optimizer that leverages the best of each optimizer. At a matched budget, omni beats every standalone optimizer. (thread)
How do we leverage the complementary performance profiles of these optimization algorithms?
Today, we release optimize_anything omni: a meta-optimizer that leverages the best of each optimizer. At a matched budget, omni beats every standalone optimizer. (thread)
2
Text Optimization is a general problem-solving paradigm, allowing one to express many real-world problems as optimizing a text artifact.
We formalized this in our optimize_anything release (demonstrated over code optimization, graphics design, algorithm and mathematical optimization, etc.), and have since seen the application of GEPA, AutoResearch, and Meta-Harness to many interesting problems.
optimize_anything’s goal is to allow users to express their problem once, and benefit from continued research into optimizers.
We formalized this in our optimize_anything release (demonstrated over code optimization, graphics design, algorithm and mathematical optimization, etc.), and have since seen the application of GEPA, AutoResearch, and Meta-Harness to many interesting problems.
optimize_anything’s goal is to allow users to express their problem once, and benefit from continued research into optimizers.
Lakshya A Agrawal@LakshyAAAgrawal
Excited to release @gepa_ai's optimize_anything: a universal API for optimizing any text parameter.
It consistently matches or outperforms domain-specific tools optimizing code, prompts, agent harnesses, cloud policies, even visuals!
If you can measure it, you can optimize it.
It consistently matches or outperforms domain-specific tools optimizing code, prompts, agent harnesses, cloud policies, even visuals!
If you can measure it, you can optimize it.
3
To understand the performance of these optimizers, we evaluated AutoResearch (program.md + coding agent, inspired by @karpathy), Meta-Harness (@yoonholeee), and GEPA (@gepa_ai) on Frontier-CS (@MangQiuyang). Several interesting phenomena show up:
Under the same model, thinking effort, and budget, no optimizer is dominating all the randomly selected tasks. Out of 10 tasks, GEPA dominates 3, AutoResearch dominates 3, and Meta-Harness dominates 4.
Under the same model, thinking effort, and budget, no optimizer is dominating all the randomly selected tasks. Out of 10 tasks, GEPA dominates 3, AutoResearch dominates 3, and Meta-Harness dominates 4.
4
We then noticed that all optimizers hill-climb very fast, making great progress with just a few iterations, but eventually come to a plateau. Switching to a different optimizer after the plateau helps unblock the progress (different optimizers approach the problem differently)!
This motivated the question: How do we leverage the complementary performance profiles of these optimization algorithms?
This motivated the question: How do we leverage the complementary performance profiles of these optimization algorithms?
5
With these insights, we build a simple meta-optimizer omni that runs all the optimizers first, with a fraction of the budget each. When all optimizers finish, we take the best candidate and hand it to a new, fresh optimizer. This meta-optimizer pipeline is only ~10 lines of code in optimize_anything!
6
The omni meta-optimizer scores +7.8 pp compared to the best standalone optimizer under the same budget, and runs faster, too!
7
To use the latest optimize_anything API, we recommend pointing your agent at gepa-ai/gepa, and asking it to use the gepa-optimize-anything skill!
Read more: gepa-ai.github.io/gepa/blog/2026…
Read more: gepa-ai.github.io/gepa/blog/2026…
8
In collaboration with an incredible team: @LakshyAAAgrawal @lukedhlee @benzhang04 Dan Klein @AlexGDimakis @koushik77 @matei_zaharia
Special thanks to @rishabhs @qiuyupeng @Mdjxjxnsk @lateinteraction for discussion and feedback on this work. We are grateful to the GEPA open-source community, too.
Special thanks to @rishabhs @qiuyupeng @Mdjxjxnsk @lateinteraction for discussion and feedback on this work. We are grateful to the GEPA open-source community, too.
