Datasets:

---

J0nasW

/

paperswithcode

like 2

I’ve been exploring 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 tools, where:

• capabilities represent atomic cognitive operations
• skills define composable workflows over capabilities
• execution is explicit and structured, instead of embedded in prompts

This tries to separate concerns that are often mixed in current systems:

• cognition (what needs to be done)
• execution (how it is carried out)
• orchestration (handled by the agent)

The hypothesis is that this separation can improve:

• composability
• observability
• control over execution

Open-source implementation:
https://github.com/gfernandf/agent-skills

Paper (DOI):
https://doi.org/10.5281/zenodo.19438943

I’d be especially interested in feedback on:

• how far capability granularity should go before overhead dominates
• whether declarative execution models can realistically replace prompt pipelines
• where this approach would break in real-world agent systems

Happy to discuss details or share specific examples if useful.