The Split between OpenAI and Microsoft, AI for Science and more.
We've developed SPL (Subsumption Prompt Language)—a framework for hierarchical LLM
agent control where state & invariants are explicit, not hidden in the LLM.
Core idea: agents maintain a symbolic world model + control logic grounded in verifiable
invariants. Result: 5-15x cost reduction, formal safety guarantees, online learning.
It applies subsumption architecture (robotics) directly to MCP-based agent orchestration.
The framework explicitly identifies open problems (confidence decay, invariant scalability,
multi-agent trust) as research directions the community should tackle.
White Paper: https://ppl-ai-code-interpreter-files.s3.amazonaws.com/web/direct-files/2aa1ae17cde839aac8432a522acc0b25/58a7c123-c15b-4e51-89a3-9eef12aff9c0/pdf_2f2cd856.pdf
Executive Pitch (5 min read): https://ppl-ai-code-interpreter-files.s3.amazonaws.com/web/direct-files/2aa1ae17cde839aac8432a522acc0b25/6c75681e-5ab8-46a1-9c3d-faff0deac14a/pdf_d3b9ea9c.pdf
Happy for feedback.
Best,
P
We've developed SPL (Subsumption Prompt Language)—a framework for hierarchical LLM
agent control where state & invariants are explicit, not hidden in the LLM.
Core idea: agents maintain a symbolic world model + control logic grounded in verifiable
invariants. Result: 5-15x cost reduction, formal safety guarantees, online learning.
It applies subsumption architecture (robotics) directly to MCP-based agent orchestration.
The framework explicitly identifies open problems (confidence decay, invariant scalability,
multi-agent trust) as research directions the community should tackle.
White Paper: https://ppl-ai-code-interpreter-files.s3.amazonaws.com/web/direct-files/2aa1ae17cde839aac8432a522acc0b25/58a7c123-c15b-4e51-89a3-9eef12aff9c0/pdf_2f2cd856.pdf
Executive Pitch (5 min read): https://ppl-ai-code-interpreter-files.s3.amazonaws.com/web/direct-files/2aa1ae17cde839aac8432a522acc0b25/6c75681e-5ab8-46a1-9c3d-faff0deac14a/pdf_d3b9ea9c.pdf
Happy for feedback.
Best,
P