Skip to content

Agent Internals

Planning Strategies

11 min+110 XPStep 1/10

ReAct's blind spot: depth-1 lookahead

ReAct plans one step at a time — which means it's greedy. It commits to the locally-best next action with no model of where the trajectory is heading, and it can't backtrack out of a dead end without you building the machinery.

That's fine — most production tasks are short-horizon and reversible. But know what you shipped: an agent that optimizes each step, not the journey. On tasks where a wrong early move forecloses the right answer (research that burns its one API call, code edits that can't be reverted), greedy fails structurally, not occasionally.