Getting started with AG-UI — the agent-user interaction protocol

While AI agents, which can perform long-running work autonomously, have supplanted simple chats for serious work with large (and small) language models, agents have a last-mile problem: How do you display what an agent is doing in front of a user and let the user act on it, without forcing every team to build its own plumbing? One answer is AG-UI, which is intended to standardize how AI agents connect to user-facing applications.

AG-UI was developed based on real-world requirements and practical experience. It grew out of CopilotKit’s initial partnership with LangChain and CrewAI, but has grown to support most of the major agent frameworks.

What are AI agents, exactly?

An AI agent is like a self-driving car for computers—it’s a software system that uses artificial intelligence to think, make plans, and do several things in a row to get what it wants. Unlike a typical chatbot that responds to a single question, an AI agent can engage in a continuous cycle of thinking, utilizing tools, analyzing outcomes, and refining its approach for the next iteration.

AI agents have a brain (usually a large language model, LLM); hands (tools that can access web browsers, APIs, or code execution); memory (short-term context in RAM, and long-term context in a database, typically a vector database); and an orchestrator (a manager for the workflow, loops, and communication between steps).

What are agent frameworks?

Agent frameworks are software toolkits that help you to build agents. Some of the most important agent frameworks are shown in the table below.

An AI agent framework is a collection of software tools, libraries, and architectural building blocks that assist developers in designing, orchestrating, and executing autonomous AI systems. An agent framework offers the infrastructure for an LLM to reason, plan, utilize tools, retain past interactions, and execute multi-step workflows, rather than simply prompting it to generate text in a single turn.

Best agent frameworks, according to LangChain

LangChain

What is AG-UI and how does it fit with MCP and A2A?

The first two widely-supported standard protocols used by agent frameworks, one to allow agents to talk to tools (MCP), and one to allow agents to talk to other agents (A2A), seemed like big deals when they were introduced, and still are, but they aren’t the end of the story. Additional protocols for agents support commerce (UCP), payments (AP2), and UI composition (A2UI). A sixth protocol to support streamed agent-user interaction (AG-UI) is the subject of this article.

MCP (Model Context Protocol) is an open-source standard framework for connecting AI applications to external systems. It was originally developed at Anthropic and then released to the Agentic AI Foundation (AAIF), under the umbrella of the Linux Foundation. MCP reduces development time and complexity when integrating an agent with data sources, tools and apps. It’s very convenient for small local projects. It’s not the most efficient or secure protocol in the world and it doesn’t really perform well at internet scale, however. If you want an efficient production integration, you’ll have to bite the bullet and use the enterprise APIs of your data sources, tools, and apps.

MCP protocol block diagram, courtesy of AAIF

AAIF

The Agent2Agent (A2A) protocol is an open standard for communication and collaboration between AI agents. It was proposed by Google in April 2025, and joined the Agentic AI Foundation (AAIF) in August 2026. A2A complements MCP, and does not replace it. With A2A, agents can discover each other’s capabilities, negotiate interaction modalities (text, forms, media), collaborate securely on long-running tasks, and operate without exposing their internal state, memory, or tools. A2A uses JSON-RPC 2.0 over HTTPS.

AG-UI is a simple, event-driven protocol to help AI agents connect to user-facing apps. It’s designed to be straightforward and adaptable, so it makes it easy for AI agents to work with the user’s interface. During agent executions, agent back ends emit events compatible with one of AG-UI’s 16 standard event types. Agent back ends can accept one of a few simple AG-UI–compatible inputs as arguments. AG-UI includes a flexible middleware layer that ensures compatibility across diverse environments. It also ships with a reference HTTP implementation and default connector.

AG-UI block diagram, emphasizing the supported agent frameworks

AG-UI

The agent protocol stack, showing the relationship among AG-UI, MCP, and A2A.

AG-UI

AG-UI architecture diagram. Note the single event stream.

AG-UI

AG-UI hands-on demos

The first time I looked at the AG-UI tools and demos in the AG-UI Dojo, I found four of them (see screenshot below). A week or so later, there were more than 30 — but the Haiku generator I had picked was no longer

[…]
Content was trimmed to protect the source. Please visit the original article for the full text.

This article has been indexed from InfoWorld