How OpenGov Replaced LangGraph with an Effect-Native Agent Loop

OpenGov abandoned LangGraph for a homegrown agent loop built on the Effect TypeScript library. The move gave them full control over tracing, structured concurrency, and model swapping—essential for complex multi-product assistants. The architectural bet paid off in operational control but demands a team comfortable with functional programming and building infrastructure in-house.

They adopted Google's Agent-to-Agent (A2A) protocol as a rigid contract between frontend and backend, ensuring that agent cards and routes are always consistent. Their core pattern is a tools-and-skills architecture: product teams independently build toolkits registered with the model, enabling parallel development of customer‑facing AI capabilities while keeping the agent loop centralized.

Safety is enforced by deterministically interrupting the loop for human approval on mutating tool calls and by running code execution inside ephemeral sandboxes. Long context is handled via rolling summarization with recall, avoiding token‑limit headaches. All of this is observed through Effect’s automatic tracing, which makes cross‑service debugging and profiling straightforward without extra instrumentation.

The talk promotes Effect heavily, but the real signal is that production agents often require custom infrastructure to meet safety, observability, and multi‑team scaling requirements. ‘Tools and skills are all you need’ is an oversimplification, yet it captures a decoupled architecture that lets model quality and business capabilities evolve independently.