Agents in Production: How OpenGov Built and Scaled OG Assist - Gabe De Mesa, OpenGov
TL;DR
Effect library bet paid off: OpenGov moved from LangGraph to a custom Effect Native Agent Loop, gaining full control over tracing, structured concurrency, and logging while maintaining the ability to hot-swap language models via dependency injection.
A2A protocol drives alignment: Adopting Google's Agent-to-Agent protocol gave them a rigorous spec (including agent cards with name and description) that served as the contract between front-end and back-end, accelerating development.
Shipping is the start, not the finish: They combine thumbs up/down user feedback with automated evals in CI to test prompts against real completions, checking if tools were hit correctly.
Safety through determinism: Humans must explicitly approve or reject tool calls that require approval, and agents execute code in ephemeral sandboxes that tear down automatically after use.
Rolling summarization beats context stuffing: Instead of stuffing all messages into context, they use running summaries after N messages plus recall over the summary, keeping only the most recent messages in full.
Agents build their own tools internally: OpenGov uses Claude and Cursor to accelerate their own development velocity while building customer-facing agent capabilities.
The Breakdown
OpenGov built OG Assist, an AI assistant integrated across their entire ERP product suite, by betting big on Effect (a TypeScript library) and building their own agent loop after outgrowing LangGraph. The talk covers their full stack approach including A2A protocol for front-end/back-end alignment, automated evals in CI, sandboxing for safe code execution, and rolling summarization for handling long conversations.
Was This Useful?
Share
Keep Reading
Make Alcreon Yours
Tune your feedFive quick questions, and the feed ranks what matters to you first.Or just get notified
The weekly Echo. Signal worth keeping in your inbox.
Every new piece, announced on X.
Read Next
See all
Playbook
The Cheapest Model That Passes
OpenRouter lists 400 models behind one API. The fix for choosing isn't a better leaderboard, it's a four-step protocol that ends in a real eval.
Playbook
Cheap Models, Hard Tasks
Most agent workflows route every step to the frontier model by default. The bill scales with how chatty the agent gets, even when most steps don't need that brain.
Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.