EvoC2F: Compiling Tool Orchestration for Efficient and Evolvable LLM Agents

Lei Wei, Qi Liu, Ruiyang Huang, Xiao Peng, Sicong Xie, Lanbo Lin, Chenhao Jiang, Yuanwu Xu, Tianyuan Yang, Jiayao Liu, Li Cai, Zhaolu Kang, Bin Wang
Proceedings of the 43rd International Conference on Machine Learning, PMLR 306:133855-133884, 2026.

Abstract

Tool-augmented language model agents have shown great potential in solving complex real-world tasks, but a key challenge remains balancing planning flexibility with the reliability required for production deployment. Existing approaches either execute tools sequentially without parallelism or generate unconstrained code, hindering optimization and verification. Additionally, agents that learn from experience often suffer from skill library pollution, where unverified abstractions degrade performance over time. We propose EvoC2F, a framework that redefines tool orchestration through program compilation and verified continuous learning. By constraining plan generation to a well-defined intermediate representation with explicit semantic annotations, EvoC2F enables provably correct optimizations, parallelism, and fault tolerance, while ensuring soundness guarantees. Our verification-gated code-to-function evolution process ensures that learned skills undergo rigorous testing before library admission. Experiments across diverse benchmarks demonstrate that EvoC2F outperforms existing methods, reducing latency and establishing a robust foundation for building reliable, evolving autonomous agents.

Cite this Paper

BibTeX
@InProceedings{pmlr-v306-wei26q, title = {{E}vo{C}2{F}: Compiling Tool Orchestration for Efficient and Evolvable {LLM} Agents}, author = {Wei, Lei and Liu, Qi and Huang, Ruiyang and Peng, Xiao and Xie, Sicong and Lin, Lanbo and Jiang, Chenhao and Xu, Yuanwu and Yang, Tianyuan and Liu, Jiayao and Cai, Li and Kang, Zhaolu and Wang, Bin}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {133855--133884}, year = {2026}, editor = {Zhang, Tong and Dudik, Miroslav and Jaggi, Martin and Agarwal, Alekh and Li, Sharon and Schuurmans, Dale and Zhu, Jerry and Berkenkamp, Felix and Dong, Hanze and Bietti, Alberto}, volume = {306}, series = {Proceedings of Machine Learning Research}, month = {06--11 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v306/main/assets/wei26q/wei26q.pdf}, url = {https://proceedings.mlr.press/v306/wei26q.html}, abstract = {Tool-augmented language model agents have shown great potential in solving complex real-world tasks, but a key challenge remains balancing planning flexibility with the reliability required for production deployment. Existing approaches either execute tools sequentially without parallelism or generate unconstrained code, hindering optimization and verification. Additionally, agents that learn from experience often suffer from skill library pollution, where unverified abstractions degrade performance over time. We propose EvoC2F, a framework that redefines tool orchestration through program compilation and verified continuous learning. By constraining plan generation to a well-defined intermediate representation with explicit semantic annotations, EvoC2F enables provably correct optimizations, parallelism, and fault tolerance, while ensuring soundness guarantees. Our verification-gated code-to-function evolution process ensures that learned skills undergo rigorous testing before library admission. Experiments across diverse benchmarks demonstrate that EvoC2F outperforms existing methods, reducing latency and establishing a robust foundation for building reliable, evolving autonomous agents.} }
Endnote
%0 Conference Paper %T EvoC2F: Compiling Tool Orchestration for Efficient and Evolvable LLM Agents %A Lei Wei %A Qi Liu %A Ruiyang Huang %A Xiao Peng %A Sicong Xie %A Lanbo Lin %A Chenhao Jiang %A Yuanwu Xu %A Tianyuan Yang %A Jiayao Liu %A Li Cai %A Zhaolu Kang %A Bin Wang %B Proceedings of the 43rd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2026 %E Tong Zhang %E Miroslav Dudik %E Martin Jaggi %E Alekh Agarwal %E Sharon Li %E Dale Schuurmans %E Jerry Zhu %E Felix Berkenkamp %E Hanze Dong %E Alberto Bietti %F pmlr-v306-wei26q %I PMLR %P 133855--133884 %U https://proceedings.mlr.press/v306/wei26q.html %V 306 %X Tool-augmented language model agents have shown great potential in solving complex real-world tasks, but a key challenge remains balancing planning flexibility with the reliability required for production deployment. Existing approaches either execute tools sequentially without parallelism or generate unconstrained code, hindering optimization and verification. Additionally, agents that learn from experience often suffer from skill library pollution, where unverified abstractions degrade performance over time. We propose EvoC2F, a framework that redefines tool orchestration through program compilation and verified continuous learning. By constraining plan generation to a well-defined intermediate representation with explicit semantic annotations, EvoC2F enables provably correct optimizations, parallelism, and fault tolerance, while ensuring soundness guarantees. Our verification-gated code-to-function evolution process ensures that learned skills undergo rigorous testing before library admission. Experiments across diverse benchmarks demonstrate that EvoC2F outperforms existing methods, reducing latency and establishing a robust foundation for building reliable, evolving autonomous agents.
APA
Wei, L., Liu, Q., Huang, R., Peng, X., Xie, S., Lin, L., Jiang, C., Xu, Y., Yang, T., Liu, J., Cai, L., Kang, Z. & Wang, B.. (2026). EvoC2F: Compiling Tool Orchestration for Efficient and Evolvable LLM Agents. Proceedings of the 43rd International Conference on Machine Learning, in Proceedings of Machine Learning Research 306:133855-133884 Available from https://proceedings.mlr.press/v306/wei26q.html.

Related Material