Effect

The Shift to Native Effect-TS Loops Building production-ready AI agents requires absolute control over execution. When the engineering team at OpenGov first launched OG Assist, an embedded AI assistant across their government ERP software, they relied on LangGraph. But scaling changed their requirements. To achieve fine-grained control, they migrated to a custom agent loop written in TypeScript using Effect. This shift allowed the team to inject different language models dynamically using clean dependency injection. Effect provides built-in schemas, error handling, and structured concurrency out of the box. By building a native loop, they gained full agency over execution, making it easier to parse tool calls and hot-swap LLMs without fighting framework abstractions. Managing Mutation with Deterministic Interrupts Safety in enterprise environments is non-negotiable, especially when agents handle municipal workflows like utility billing or asset management. To prevent unauthorized database changes, the platform implements strict boundaries. When an agent triggers a mutating tool call, the system deterministically interrupts the run. It pauses the execution thread and renders a dedicated approval UI. The user must explicitly accept or reject the action. For broader sandboxing, any code execution or file creation occurs in isolated, ephemeral environments that tear down automatically, ensuring complete isolation from production systems. Tackling Context Bloat with Rolling Summaries Long-running conversations quickly degrade model performance and break token limits. Stuffing historical messages into the prompt is a recipe for failure. To solve this, the team implements a rolling summarization strategy. After a set number of turns, the system summarizes the conversation history up to that point. It retains only the most recent messages in raw format, while using the running summary for memory recall. If a user refers to an event from a hundred turns prior, the agent retrieves the context from the summary, preserving accuracy without inflating latency. Native Tracing Eliminates Production Blind Spots You cannot scale what you can't see. Debugging multi-step agent behaviors across microservices is notoriously difficult. By building on top of the Effect ecosystem, the team gets distributed tracing out of the box. Every functional span is tagged automatically. When a tool call slows down or fails, developers can inspect the exact execution path, isolate bottlenecks, and cross-reference data across services. Combining these traces with real-time feedback loops—like automated testing in CI and user-driven thumbs-up metrics—allows the engineering team to deploy updates with confidence.