Optimizing Agents

The library behind our agent-optimization engagements, out of stealth: a PyTorch-shaped training loop for agent runtimes, with textual gradients, batch-level updates, and constraint-preserving steps. Why we built it instead of adopting what existed.