Gabe De Mesa is a software engineer who helped build OG Assist, an AI assistant embedded in government software. His approach to production agents combines application-specific tools, human approval, isolated execution, and detailed operational monitoring so public-sector employees can use AI without losing control of consequential decisions.
De Mesa earned a bachelor’s degree in computer science from the University of California, Irvine, in 2021 and a master’s degree in computer science from the University of Illinois Urbana-Champaign in 2024. His earlier engineering work included Bayer’s ForGround agricultural platform and Climate FieldView. At OpenGov, he worked on utility-billing software before joining its AI agents team, where he helped integrate OG Assist into government workflows spanning finance, procurement, and other operational tasks.
How he builds agents that can actually ship
- An Effect-native agent loop: De Mesa and his team replaced an initial LangGraph implementation with an agent loop built using Effect and TypeScript. Owning the orchestration layer gave them finer control over structured concurrency, error handling, logging, tracing, and interchangeable language models.
- Agent2Agent protocol as shared infrastructure: Agent descriptions, message schemas, and routes establish a common contract across frontend and backend systems. Registered tools connect the assistant to product-specific capabilities, while reusable interface components let it generate relevant forms and choices dynamically.
- Human-in-the-loop approval: Potentially consequential actions interrupt the agent loop until someone explicitly approves or rejects them. Code execution and generated files run inside temporary, isolated sandboxes that are removed after use.
- Rolling conversation summaries: Combining compressed historical context with recent messages allows agents to recall earlier exchanges without repeatedly passing an entire conversation into the model. Automated evaluations test whether the assistant selects appropriate tools, customer feedback reveals weak responses, and distributed traces expose bottlenecks and failures.
De Mesa also applies agent tooling internally, using Claude, Cursor, and custom tools to support code development and review. His AI Engineer World’s Fair presentation makes his central technical priority clear: agents become useful in sensitive environments when their tools, permissions, memory, and behavior remain legible to the people operating them.
