CopilotKit ships AG‑UI, AIMock and Pathfinder to close production gaps for agentic AI

CopilotKit said it has completed a 2026 shipping cycle that delivers three infrastructure pieces it calls essential to moving agentic AI from demos into production: AG‑UI (an interaction protocol), AIMock (a deterministic testing suite released in April 2026) and Pathfinder (a server focused on runtime persistence). The company, co‑founded by Atai Barkai and Uli Barkai and based in Seattle, frames the work as closing the practical gaps that separate prototype agents from production‑grade systems.

AG‑UI defines the interaction layer between agents, applications and end users instead of treating agents as passive chat widgets. The protocol supports real‑time streaming responses, dynamic UI component generation, bidirectional state synchronization between agent and app, and human‑in‑the‑loop pauses where agents wait for user confirmation before proceeding. CopilotKit lists major cloud vendors as supporters — Google, Microsoft, Amazon and Oracle — and cites integrations with frameworks such as LangChain, Mastra, PydanticAI and Agno.

First‑party SDKs for AG‑UI are available for LangGraph, CrewAI, Mastra, Agno and Pydantic AI, and community implementations now cover Kotlin, Go, Dart, Java, Rust, Ruby and C++. Additional ports for.NET, Nim, Flowise and Langflow are in progress, reflecting a push for broad runtime and language support.

AIMock, released in April 2026, targets a persistent testing problem: a single agent request commonly touches six or seven live services — typically an LLM, an MCP tool server, a vector database, a reranker, a web‑search API, a moderation layer and sometimes sub‑agents communicating over A2A. Teams that mock only one dependency produce non‑deterministic test suites; AIMock aims to replace that fragility with a single, centralized approach—“one JSON config, one port”—that can mock every dependent service for repeatable tests.

The AIMock configuration reportedly covers eleven LLM providers, with explicit support mentioned for OpenAI, Anthropic Claude, Google Gemini and Amazon Bedrock/Azure endpoints. By centralizing mocks for LLMs, tool servers and other integrations, CopilotKit positions AIMock as a way to make continuous integration and local testing yield provable, repeatable behavior for end‑to‑end agent flows.

Pathfinder, the third piece of the cycle, is described as a server focused on runtime persistence to help agents maintain state and continuity across sessions — another practical requirement for production deployments. CopilotKit positions AG‑UI alongside protocol work for tool access (MCP) and agent‑to‑agent coordination (A2A), arguing the three operate at different stack layers similar to TCP, HTTP and HTML. As ecosystem signals, the company notes that AWS has integrated AG‑UI into FAST examples and Bedrock AgentCore, and that Atai Barkai is teaching a full‑stack AG‑UI course on DeepLearning.AI pairing a LangChain backend with a React frontend and the AG‑UI runtime.