ORCA: A Cognitive Runtime Layer for Agent Systems (paper + open source)

I’ve been exploring some of the structural limitations of prompt-based agent systems and built a framework to experiment with an alternative approach.

The core idea is to introduce a “cognitive runtime layer” (ORCA) between the agent and the underlying tools.

In this model:

• capabilities represent atomic cognitive operations (e.g., retrieve, transform, evaluate)
• skills define composable workflows over capabilities
• execution is explicit and structured, rather than embedded in prompts

This aims to separate concerns that are often tightly coupled in current agent designs:

• cognition (what needs to be done)
• execution (how it is carried out)
• orchestration (decision-making at the agent level)

The hypothesis is that making this separation explicit can improve:

• composability
• observability
• controllability of execution

Open-source implementation:

github.com

GitHub - gfernandf/agent-skills: Agents should execute whenever possible — runtime...

Agents should execute whenever possible — runtime for composable AI agent skills

Paper (DOI):

Zenodo

Beyond Prompting: Decoupling Cognition from Execution in LLM-based Agents...

Recent advances in large language model (LLM) agents have largely relied on prompt-centricdesigns, where complex tasks are executed through monolithic, single-shot or loosely structuredprompting strategies. While effective in some settings, this...

I’d be particularly interested in feedback on a few points:

• how far capability granularity should go before overhead dominates
• whether declarative execution models can realistically replace prompt pipelines
• how this kind of abstraction would behave in more complex, real-world agent systems

Happy to expand on the execution model, design principles, or concrete examples if useful.