Deep dives on neurosymbolic agents, open-source runtime architecture, medical administration, typed execution, and high-trust AI systems.
How runtime subtypes and shadow discovery let the ontology reason about care-management eligibility, APCM code selection, and payer-specific patterns.
A live proof that guarded runtime actions can stop an agent from skipping review, self-approving, or mutating state out of order.
How an executable graph breaks the static toolbox by creating, activating, and using new capabilities while work is still in progress.
A technical explanation of Buffaly's second supervisory loop: how a separate model watches the working agent's memory, catches drift, preserves continuity through compaction, and sends the worker back to concrete tool work when it matters.
A technical explanation of executable graph agents: how semantic identity, typed objects, runtime actions, native code, and self-extending capability change what agents can learn and execute.
A practical evaluation of local embedding models for Buffaly's short action/entity semantic retrieval workload, including methodology changes, run IDs, EmbeddingIDs, storage caveats, and reproducibility notes.
How language acquisition, dual-channel learning, ontology, ProtoScript, and executable memory shaped the long path to Buffaly.
A case for building runtime-first systems around frontier models instead of asking larger prompts to become memory, execution, policy, and control.
A different kind of agent: one that turns language into executable structure instead of keeping everything in text prompts.
Why traditional LLM agents are an operational dead end in medical administration: and why we built a neurosymbolic alternative.
