EvoC2F: Compiling Tool Orchestration for Efficient and Evolvable LLM Agents

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.

AI systems are increasingly able to use external tools, such as search engines, databases, apps, and online services, to complete real-world tasks. However, these systems can be slow or unreliable when they call tools one by one, or unsafe when they generate free-form code that is hard to check. This paper introduces EvoC2F, a framework that helps such AI systems plan tool use in a clearer and safer way. EvoC2F breaks a task into well-defined steps, records which steps depend on each other, and identifies which steps may affect shared resources or the outside world. This allows the system to run independent steps in parallel while keeping risky actions in the right order. EvoC2F also learns reusable routines from past successful tasks, but only adds them after testing that they work and do not hurt performance on other tasks. Experiments show that this approach improves success rates, reduces waiting time, and supports more reliable long-term learning for tool-using AI agents.