Design, Potential and Realization in Agentic Systems
A framework for agentic systems that connects configuration to possible and realized behaviors.
Abstract
Existing work usually treats model and tool selection, access control, resource allocation, workflow composition, and evaluation as separate design problems. This separation leaves no common object relating a system’s complete pre-execution specification to both its permitted and realized behavior. We introduce a typed Design object that combines models, tools, data access, authority, resources, composition, and control policies. For each task, its contracts define Potential as the set of admitted traces, while its models and policies induce a probability law over realized traces. The paper makes four linked contributions:
- First, we give typed execution semantics in which compatible component traces compose invariantly under the declared coherence conditions, and executed events remain within previously granted authority under the enforcement and transfer assumptions.
- Second, we define reliability-constrained selection using Potential-minimality and Pareto efficiency. The former rules out a reliable Design when another reliable Design admits strictly fewer traces, whereas the latter rules out a reliable Design dominated across cost, latency, exposure, risk, and overhead. Neither criterion implies the other.
- Third, we establish that passive traces cannot identify a permission-minimal reliable repair when unexercised permission extensions induce the same observed law, which makes support necessary for logged evaluation.
- Fourth, for direct evaluation over a finite safe catalog, we derive simultaneous confidence intervals and a sound classification rule that removes Designs shown to be unreliable or dominated while retaining unresolved candidates.