Summary
Gabe de Mesa, a software engineer on the AI Agents team at OpenGov, delivers a production-focused conference talk on building and scaling OG Assist — an AI agent embedded across OpenGov’s government ERP software covering budgeting, procurement, asset management, and permitting for public-sector clients. OG Assist surfaces as a persistent chat interface across every product, letting users query context-specific data and trigger on-screen actions through natural language, with each product team contributing purpose-built tools and skills.
A central architectural decision was adopting Effect, an open-source TypeScript library, as the backbone of a fully custom agent loop rather than relying on off-the-shelf frameworks. This gave the team fine-grained control over dependency injection, model swapping, logging, distributed tracing, and error handling. Google’s Agent-to-Agent (A2A) protocol was adopted to define typed contracts between frontend and backend agents, improving alignment across teams and enabling extensible agent routing. The A2A agent card spec is used to formalize each agent’s name, capabilities, and interface.
De Mesa walks through the team’s complete production feedback cycle: thumbs-up/thumbs-down signals captured from users, automated CI evals that run real completions and verify expected tool calls, and a deterministic human-in-the-loop interrupt mechanism that pauses the agent loop and surfaces an approval UI whenever a high-stakes tool call is about to execute. Long-context management, sandboxing, and trace-based observability round out what is a thorough real-world blueprint for enterprise agent deployment.
📺 Source: AI Engineer · Published June 26, 2026
🏷️ Format: Hands On Build
