Combee: Scaling Prompt Learning for Self-Improving Agents (April 2026)

Free

Berkeley/Stanford (Stoica, Zou, Gonzalez): scales parallel prompt learning with up to 17x speedup over ACE/GEPA via parallel scans and dynamic batching; evaluated on AppWorld, Terminal-Bench, FiNER

FreeFree tier

Type

Open Source

Visit Website GitHub (40 stars)

About Combee: Scaling Prompt Learning for Self-Improving Agents (April 2026)

Combee is a novel framework from UC Berkeley and Stanford researchers (including Stoica, Zou, Gonzalez) that scales parallel prompt learning for self-improving language model agents. It addresses the quality degradation that occurs when existing methods like ACE and GEPA are applied with high parallelism. Combee leverages parallel scans and an augmented shuffle mechanism to learn from many agentic traces simultaneously, and introduces a dynamic batch size controller to balance quality and delay. Evaluated on AppWorld, Terminal-Bench, Formula, and FiNER, Combee achieves up to 17x speedup over previous methods while maintaining comparable or better accuracy and equivalent cost.

Key Features

Parallel prompt learning across multiple agentic traces without quality degradation

Parallel scans and augmented shuffle mechanism for efficient learning

Dynamic batch size controller to balance quality and delay

Compatible with existing prompt learning methods like ACE and GEPA

Up to 17x speedup over prior methods with comparable accuracy

Pros & Cons

Pros

Significant speedup (up to 17x) for prompt learning without sacrificing quality
Maintains or improves accuracy compared to sequential methods
Enables efficient learning from many parallel agent traces
Dynamic batching adapts to workload to balance performance and latency

Best For

Improving LLM agent accuracy by learning task-relevant prompts from inference-time contextScaling prompt learning from large collections of agentic tracesParallel agent execution with aggregate trace learningSelf-improving agents that refine their prompts based on past runs

FAQ

What problem does Combee solve?

Existing prompt learning methods like ACE and GEPA suffer from quality degradation when scaled to high parallelism. Combee enables efficient parallel prompt learning without quality loss.

How does Combee achieve speedup?

Combee uses parallel scans and an augmented shuffle mechanism to process multiple agentic traces concurrently, along with a dynamic batch size controller to balance quality and delay.

What benchmarks was Combee evaluated on?

Combee was evaluated on AppWorld, Terminal-Bench, Formula, and FiNER, achieving up to 17x speedup over previous methods with comparable or better accuracy and equivalent cost.