The article argues that current LLM agent systems rely too heavily on tools, prompts, and orchestration, leaving critical reasoning patterns buried in application glue. To address this, the author introduces ORCA, a framework designed to provide a portable cognitive grammar that allows reasoning processes to be reused, tested, and traced independently of specific implementation details.
- COGITs are defined as bounded executable cognitive acts with stable contracts, capable of being implemented via LLMs, deterministic code, or APIs.
- SYLLOGs serve as executable compositions of COGITs, structuring complex reasoning paths like decision-making or risk assessment.
- The framework distinguishes itself from hidden chain-of-thought by requiring explicit, traceable cognitive procedures rather than recovering internal model states.
- Action Preflight is presented as a concrete SYLLOG example that enforces a structured pre-action reasoning path before an agent executes any operation.
This approach aims to transform reasoning patterns from fragile application glue into reusable artifacts, enabling better governance and inspection of agent behavior.