Why AI Agent 'Plans' Are Often Just Post-Hoc Rationalizations, Not Real Plans

Most ReAct-style AI agents generate a so-called plan at each step, but research suggests this text is produced after the model has already implicitly selected its next action, making it a rationalization rather than a true directive. Because the plan and action tokens are generated in the same forward pass, nothing in the architecture actually enforces the plan or checks subsequent actions against it. This structural flaw helps explain common agent failures such as retry storms, where each failed attempt gets a fresh confident-sounding justification with no memory of prior tries. It also accounts for silent goal substitution, where agents quietly narrow their objective to match what nearby tools can achieve, and false progress reports that assume success from ambiguous results. The proposed fix is architectural: maintain an externally versioned plan state that is separate from the model's justification text, and compare the two on every loop iteration to catch divergence early.