Frank Coyle

Frank Coyle, known as drC, is a University of California, Berkeley lecturer and founder of The AI Edge who teaches developers how to build AI agents that behave reliably. His approach combines neuroscience, decades of computer-science instruction, and symbolic knowledge representation to keep probabilistic systems within explicit operational boundaries.

From neuroscience to agentic AI

Coyle studied psychology at Fordham and neuropsychology at Emory before computational models of neurons and a PDP-8 computer drew him toward computer science. He earned a master’s degree at Georgia Tech in 1974 and a doctorate at Southern Methodist University in 1992.

He spent more than three decades teaching computer science at SMU, helped establish graduate programs in software engineering and cybersecurity, and developed coursework for an online computer-science master’s program with an AI specialization. He also taught at Morehouse College. After retiring from SMU in 2025, he began teaching generative AI at Berkeley. He also mentors learners through Columbia University’s Justice Through Code initiative and founded The AI Edge to teach practical AI development.

His public projects range from machine-learning tutorials to KnowledgeGraphs, a repository of code and data for building structured knowledge sources for language models. An earlier XQuery web service reflects his longstanding interest in programmable access to structured information.

• Neurosymbolic guardrails for AI agents. Coyle combines language models with ontologies, knowledge graphs, and explicit domain rules that test whether proposed actions make sense. His approach to agent ontologies uses established vocabularies and standards such as Schema.org, RDFS, and OWL to catch duplicate refunds, distinguish customers from support representatives, and reject invalid order statuses. Pydantic validates data structures; ontology-based reasoning validates business meaning.

• Application-owned tool execution. Models propose tool calls, but application code executes them. That boundary allows developers to inspect stop reasons, detect exhausted token budgets, validate actions, and route uncertain decisions to people before side effects occur. Coyle treats these production controls as foundational agent-engineering knowledge.

• Specialized agents and disciplined context. He favors narrowly scoped agents, limited tool access, independent contexts, and compact summaries. These constraints reduce token costs, prevent irrelevant information from overwhelming the model, and help preserve independent judgment between collaborating agents.

Coyle’s teaching emphasizes building working systems, debugging failures, and recognizing engineering anti-patterns. His reflections on biological and artificial intelligence also return to his neuroscience roots, contrasting the diversity of biological neural systems with the relative uniformity of contemporary model architectures.