Roy Derks is a Dutch software developer, author, and IBM developer-experience specialist focused on making AI agents more reliable through better tools. His work applies lessons from GraphQL and web-application architecture to a central challenge of agent engineering: connecting language models to external systems without surrendering control over how those systems operate.
Derks taught himself programming at 14 while trying to build a DVD-exchange platform. He later founded and worked at startups, became a developer advocate, and wrote React Projects. At the GraphQL company StepZen, he worked on API-focused developer education before IBM acquired the company in 2023. Through Hackteam, he continues teaching web technologies and AI.
How he approaches agent infrastructure
- Tools as application architecture. Derks treats tool failures as core engineering problems: models select inappropriate operations, generate malformed arguments, encounter unreliable services, or return unusable data. He advocates descriptive tool definitions, typed inputs, explicit output schemas, and structured responses that make multistep workflows more predictable.
- Framework-independent agent tools. He separates tool creation and execution from orchestration so capabilities can move between LangChain, LangGraph, CrewAI, and AutoGen. His work on IBM’s watsonx.ai Flows Engine illustrates how remotely hosted tools can centralize execution, authorization, and error handling while remaining accessible across agent frameworks. His LangGraph implementation guide demonstrates that architecture in practice.
- Dynamic GraphQL and SQL tools. Instead of defining separate operations for every customer, order, or database query, Derks explores giving models existing schemas and having them generate appropriate queries. This preserves established business logic and reduces tool proliferation, but deeply nested schemas, unusual data types, invalid syntax, and hallucinations can compromise reliability.
- Model Context Protocol interoperability. He sees MCP as a practical boundary between agent clients and tool-serving infrastructure, keeping credentials, execution, and service-specific logic closer to the systems they govern.
Derks also regards AI-assisted programming as a way to recover the excitement of learning to code: faster experimentation, shorter debugging cycles, and more immediate feedback when building software.
