Decision Layer
An Agent control-flow architecture pattern that replaces regex-based schemes for probabilistically parsing LLM output with typed, probabilistic decision primitives. It consists of three parts: typed decision calls, threshold policies, and audit logs.
Timeline (last 90 days)
The system splits the Agent architecture into a memory layer (recording structured experience: tasks, actions, results and verification evidence) and a decision layer (providing typed probabilistic decisions), with the boundary between the two layers being: memory answers "is this true", and decision answers "what should we do and how certain are we".
The decision layer backend adopts a Bayesian Beta posterior, updating based on historical results in memory, and confidence moves as evidence accumulates.
The decision layer provides typed probabilistic decisions, each decision carrying its own uncertainty and escalation flag, and at runtime it can choose to hand off rather than guess when uncertain.
The decision layer pattern consists of three parts: typed decision calls (predicate/choice/score), threshold policy, and audit logs.
The decision-making layer is responsible for planning steps, decomposing tasks, and scheduling execution order.
Separating the perception layer from the decision layer provides an auditability advantage, making it possible to determine whether a problem originated in the perception layer or the decision layer when errors occur.
All Facts (6)
The system splits the Agent architecture into a memory layer (recording structured experience: tasks, actions, results and verification evidence) and a decision layer (providing typed probabilistic decisions), with the boundary between the two layers being: memory answers "is this true", and decision answers "what should we do and how certain are we".
50%UnverifiedThe decision layer backend adopts a Bayesian Beta posterior, updating based on historical results in memory, and confidence moves as evidence accumulates.
50%UnverifiedThe decision layer provides typed probabilistic decisions, each decision carrying its own uncertainty and escalation flag, and at runtime it can choose to hand off rather than guess when uncertain.
50%UnverifiedThe decision layer pattern consists of three parts: typed decision calls (predicate/choice/score), threshold policy, and audit logs.
50%UnverifiedThe decision-making layer is responsible for planning steps, decomposing tasks, and scheduling execution order.
50%UnverifiedSeparating the perception layer from the decision layer provides an auditability advantage, making it possible to determine whether a problem originated in the perception layer or the decision layer when errors occur.
50%