Fight context rot with garbage collection

AI agents suffer from 'context rot,' where performance degrades as the context window fills with old, irrelevant information over many reasoning steps. Current methods like summarizing and compressing the context are lossy and inefficient, often discarding important details. A better approach is modeled on garbage collection in traditional programming, where context is pruned at logical boundaries, similar to how a function discards local variables after execution. This 'prune-as-you-go' technique keeps the agent focused, minimizes noise, and enables more reliable and efficient performance for complex, long-running autonomous tasks.