GitHub - eminsk/agentjit: Just-In-Time Compiler for AI Agent Trajectories. Compiles multi-step LLM workflows into 0.08ms deterministic Python code with zero token cost.

AgentJIT traces dynamic AI agent trajectories and compiles them into pure, type-safe Python code, analogous to how V8 compiles hot JavaScript and PyTorch's torch.compile traces tensors into CUDA kernels. The project claims up to 100,000x speedup on hot paths and zero token cost for compiled workflows. Runtime guards trigger de-optimization, falling back to the LLM agent on unexpected inputs. The tool is framework-agnostic and installable via pip.