Beyond Prompting: Decoupling Cognition from Execution in LLM-based Agents through the ORCA Framework

Recent advances in large language model (LLM) agents have largely relied on prompt-centric
designs, where complex tasks are executed through monolithic, single-shot or loosely structured
prompting strategies. While effective in some settings, this approach results in implicit and
opaque forms of cognition that hinder composability, observability, reuse, and governance.
In this work, we argue for a principled separation between cognition and execution in LLMbased
agents. We introduce ORCA (Open Cognitive Runtime for Agents), a conceptual
and architectural framework that models cognition as the composition of explicit semantic units,
termed capabilities, orchestrated through declarative workflows, or skills, and executed via a
decoupled runtime layer. This separation enables structured reasoning processes, explicit control
over execution, and dynamic binding to heterogeneous implementations, including LLM-based
and non-LLM tools.
We formalize the core primitives of the ORCA framework and describe its architectural
realization, including a capability registry, a workflow execution engine, and a flexible binding
mechanism. We further introduce a governance model that incorporates trust levels, validation
gates, and execution traceability, enabling safer and more interpretable agent behavior.
Through comparative experiments, we analyze the trade-offs between prompt-based approaches
and structured cognitive execution, showing that while ORCA introduces additional
computational overhead, it significantly improves composability, transparency, and controllability.
We discuss the implications of this trade-off and outline scenarios where structured cognitive
runtimes provide clear advantages.
Overall, this work positions cognitive runtime layers as a promising abstraction for structuring
and governing agent cognition, particularly in settings where control, transparency, and
composability are required.