When working with AI, what costs more than a wrong answer is believing it and building on top of it. By the time you notice, going back is rarely easy.
Cogitors is an open-source Agent Skills project I built to avoid leaving important decisions to a single model.
It brings Codex, Claude and Antigravity to the same table. All three receive the same task: evaluate an architecture proposal, review a code change or find the weak points in a product decision.
The process runs over five rounds:
Independent analysis: Each agent works without seeing the others’ views.
Reconsideration: Views are shared, and agents can update their assessments.
Debate: Each agent supports its position with evidence and challenges points it disagrees with.
Final position: Objections are answered, and each agent settles on its final stance.
Synthesis: The agent that started the session — the Elder — brings the assessments together into one reasoned decision.
The core principle: All three agreeing does not mean all three are right. Reasoned minority views and unresolved uncertainties are preserved in the final decision.
The package includes four skills: /cogitors for the joint deliberation, and /codex, /claude and /antigravity for assigning tasks to each agent individually. Not every task needs a debate.
Cogitors takes its name from the minds in the Dune universe that leave their bodies behind and devote themselves to thought.
Open source under the MIT license. Installation and usage: https://github.com/erayendes/cogitors/
If Cogitors helped you catch a flawed assumption earlier or examine an important decision more carefully, that is exactly what it is for.
Your support helps me keep building these tools. Thank you.