Abstract
The central challenge for AI agents is not performance but accountability. An agent that produces correct outputs through an opaque process offers no basis for verifying why a decision was permitted, attributing errors to identifiable causes, or delegating authority with justified confidence. We introduce the Structured Cognitive Loop (SCL), an architecture that separates cognition, control, action, and memory: the language model proposes, a lightweight controller mediates every proposal before action, and memory is kept externally as a structured store of verified facts. Against ReAct and common LangChain agents on three tasks under matched models, tools, and decoding, SCL achieves 86.3% task success versus 70.5-76.8% across 360 episodes, with higher goal fidelity, fewer redundant calls, more reliable reuse of intermediate states, and fewer unsupported assertions; ablations show that control and external memory contribute independently. We then situate SCL within a broader accountability architecture comprising context-aware human-in-the-loop control, Pool-Gated Retrieval, and the Horizon-Warrant-Commitment framework, and outline the current direction of development, in which instructions are compiled into programs validated before execution so that execution authority derives from a verified program rather than per-cycle model interpretation. For the kind of accountability it pursues, SCL adopts neither multi-agent consensus as a proxy for truth nor the imitation of human cognition as a design goal: it institutionalizes cognition rather than imitating it.