跪拜 Guibai
← All articles
Artificial Intelligence · Open Source

AG-UI 1.0: A Frozen, Open Protocol For Streaming AI Agents to Any Frontend

By 冬奇Lab ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Frontend teams can stop writing a custom adapter for every agent framework; one typed event stream handles streaming text, tool calls, state, and human approval across web, mobile, terminal, and Slack. AG-UI's freeze at 1.0 gives a stable target, but teams should check how deeply their chosen framework is integrated before committing.

Summary

AG-UI is an open, MIT-licensed protocol that standardizes how AI agents talk to user-facing applications. Instead of a single request-response cycle, agents emit a typed event stream: streaming text, tool calls, shared state updates, reasoning traces, interrupts for human approval, and subagent runs. Any compatible frontend can render those events, so one agent works across the web, mobile, a terminal, or Slack without per-framework adapters. The protocol sits alongside MCP and A2A: MCP provides tools, A2A handles agent-to-agent, and AG-UI covers agent-to-user.

AG-UI 1.0 is frozen with JSON Schema support and backward compatibility between 0.x and 1.0. It defines 31 event types across eight families (though a wildcard in the reasoning family means the README's count includes expansions not individually listed). Transport is HTTP+SSE or HTTP+Protobuf. State sync uses snapshot and delta events with JSON Patch for partial updates. Human-in-the-loop ends a run with an interrupt event; the next run resumes from there, so servers don't hold long-lived connections. Official SDKs cover TypeScript, Python, and .NET, with community SDKs for Kotlin, Go, Java, Dart, Rust, Ruby, C++, and Swift. The project is initiated by CopilotKit and reports adoption by Google, Microsoft, Amazon, and Oracle, with integrations for LangChain, CrewAI, and many other frameworks.

The main caveats: not every framework has native integration, and ready-to-use frontend components lean toward CopilotKit's ecosystem. The spec is frozen, so new features wait for draft specs. This analysis is based on the README and repo structure, not a live run.

Takeaways
— AG-UI 1.0 is a frozen, MIT-licensed spec defining 31 event types across eight families, with a JSON Schema backing it.
— The protocol uses HTTP+SSE or HTTP+Protobuf as its two standard transports; other transports require custom middleware.
— Shared state syncs via STATE_SNAPSHOT and STATE_DELTA, where deltas use JSON Patch to avoid re-sending full state.
— Human-in-the-loop stops a run with an interrupt event and resumes in a subsequent run, avoiding long-held server connections.
— Official SDKs exist for TypeScript, Python, and .NET; community SDKs cover Kotlin, Go, Java, Dart, Rust, Ruby, C++, and Swift.
— AG-UI is positioned as complementary to MCP (tools) and A2A (agent-agent), covering the agent-to-user leg.
— The project is created by CopilotKit and reports adoption by Google, Microsoft, Amazon, and Oracle, with integrations for LangChain, CrewAI, and other frameworks.
Conclusions

Ending a run with interrupt instead of holding a connection is a pragmatic design that lets servers free resources between turns, at the cost of re-sending context on resume.

Because AG-UI is initiated by CopilotKit, the protocol is open but the ready-made frontend components lean on CopilotKit's ecosystem, so teams should weigh that when choosing a stack.

The spec's freeze at 1.0 gives stability but pushes new features to draft specs; adoption depth varies across frameworks, as the README itself notes via middleware recommendations.

The event model's eight families cover everything from token streaming to subagents, so a frontend can treat any agent framework as a black box that emits a known event sequence.

Concepts & terms
MCP
Model Context Protocol, a standard that gives AI agents access to tools and context; in AG-UI's trio, MCP covers the agent's tooling side.
A2A
Agent-to-Agent protocol for enabling collaboration between AI agents; alongside MCP and AG-UI, it handles the agent-to-agent leg.
JSON Patch
A format for expressing partial updates to JSON documents; AG-UI uses it in STATE_DELTA events to sync shared state without sending full snapshots each time.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗