What if OpenAI Agent Builder could do something it doesn’t natively support yet?
Reusable, auditable cognition.
I built a small ORCA-powered agent on top of OpenAI Agent Builder + MCP that can take a strategic decision request like:
“Should we launch a legal SaaS in Spain during the next 12 months?”
and instead of producing a black-box answer, it executes an explicit cognitive workflow:
prompt → routing → reusable decision skill → execution trace → confidence → report
The interesting part is not the recommendation.
The interesting part is that the reasoning becomes:
reusable
traceable
auditable
composable
And the output stops looking like:
“here are some thoughts…”
and starts looking closer to something a real team could actually use:
-
explicit recommendation
-
alternatives evaluated and scored
-
confidence levels
-
decision quality assessment
-
uncertainties and missing information
-
execution diagnostics
-
cognitive trace of what happened
-
graceful degradation when execution fails
Under the hood, the agent delegates cognition to an ORCA skill exposed through MCP.
The surprising thing?
It is actually much more practical than I expected to expose a reusable skill via MCP and plug it into Agent Builder.
But it already feels surprisingly close to something agent platforms will eventually need.
Because once agents become business-critical:
“trust me bro, the model reasoned” probably won’t be enough.
This is part of ORCA (Open Cognitive Runtime for Agents) — an open-source framework in progress for reusable cognition in agents.
Repo:
https://github.com/gfernandf/agent-skills
Paper:
https://zenodo.org/records/19438943
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6600840
Curious question for people building agents:
Should cognition stay inside prompts, or should it become a reusable runtime layer?