📄 Research Paper¶

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

Guillermo E. Fernandez Alvarez · 2026

Also available on SSRN: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6600840

Abstract¶

Modern AI agents are predominantly built around prompt-driven execution, where reasoning remains implicit inside the model and tool orchestration lacks formal structure. This paper introduces ORCA (Open Cognitive Runtime Architecture), a framework that decouples cognitive processes from execution mechanics in LLM-based agents. ORCA defines a Cognitive Execution Layer where agent behavior is expressed through composable capability contracts, declarative skill graphs (DAGs), and explicit cognitive state — replacing ad-hoc prompting with structured, inspectable, and governable execution.

The paper presents Agent Skills Runtime as a reference implementation of ORCA, demonstrating that deterministic execution of agent workflows is achievable without sacrificing flexibility. The architecture supports four binding protocols (PythonCall, OpenAPI, MCP, OpenRPC), ships 122 governed capabilities with zero-config Python baselines, and provides formal alignment with the CoALA cognitive architecture framework.

Key contributions include: (1) a formal separation of intent, cognition, and execution in agent systems; (2) a contract-first capability model with multi-protocol binding resolution; (3) a typed cognitive state (CognitiveState v1) aligned with CoALA; and (4) empirical evidence that structured execution outperforms prompt-driven approaches in reproducibility, safety, and observability.

Key Contributions¶

#	Contribution	Impact
1	Cognition–Execution Separation	Agents reason through explicit state, not implicit prompts
2	Contract-First Capabilities	Backend-agnostic operations with multi-protocol binding
3	CognitiveState v1	Typed Frame/Working/Output/Trace aligned with CoALA
4	Safety-First Execution	4-tier trust model with validation gates and scope constraints
5	Empirical Evaluation	Reproducibility, safety, and observability benchmarks

Download¶

📥 Download PDF

📄 SSRN Entry

Citation¶

If you use ORCA or Agent Skills Runtime in your research, please cite:

APA 7¶

Fernandez Alvarez, G. E. (2026). Beyond Prompting: Decoupling Cognition from Execution in LLM-based Agents through the ORCA Framework. Zenodo. https://doi.org/10.5281/zenodo.19438943

BibTeX¶

@article{fernandez_orca_2026,
  author    = {Fernandez Alvarez, Guillermo E.},
  title     = {Beyond Prompting: Decoupling Cognition from Execution in
               LLM-based Agents through the ORCA Framework},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.19438943},
  url       = {https://doi.org/10.5281/zenodo.19438943}
}

Chicago¶

Fernandez Alvarez, Guillermo E. 2026. "Beyond Prompting: Decoupling Cognition from Execution in LLM-based Agents through the ORCA Framework." Zenodo. https://doi.org/10.5281/zenodo.19438943.

Resource	Description
ORCA Specification	The full ORCA architecture spec
CoALA Mapping	Formal alignment with CoALA cognitive architecture
Positioning	Comparison with SCL, SPIRAL, CoALA, and other frameworks
Agent Skills Runtime	Reference implementation
Agent Skill Registry	Capability & skill registry

How to Cite the Software¶

For citing the runtime software itself (not the paper), use:

@software{fernandez_agent_skills_2026,
  author    = {Fernandez Alvarez, Guillermo},
  title     = {Agent Skills Runtime},
  year      = {2026},
  url       = {https://github.com/gfernandf/agent-skills},
  version   = {0.1.0},
  license   = {Apache-2.0}
}

GitHub also provides a "Cite this repository" button powered by CITATION.cff.