Separate agent planning from execution: agents output intentions, a deterministic system executes them and logs everything, preventing state loss a...
This paper solves a critical problem with AI agents: they lose track of what they're doing over long tasks and can't reliably execute code changes. ESAA is an architecture that separates what an agent *intends* to do from what actually *happens* in your codebase.