AG-UI 1.0 freezes the agent UI protocol: stop writing glue
AG-UI 1.0 is a frozen, open protocol for connecting AI agents to app frontends, with interrupts, subagents and token usage. Build your agent UI on it.
The protocol between your AI agent and your app's screen just stopped moving. On September 30, CopilotKit released AG-UI 1.0, a stable spec for how an agent backend streams events to a user-facing frontend. MCP connects agents to tools. AG-UI connects agents to people. Until now, most teams wrote that second layer by hand, once per agent framework. AG-UI 1.0 is the reason to stop.
What actually happened
Per CopilotKit's announcement, AG-UI 1.0 ships a full specification and a JSON Schema. The TypeScript, Python and .NET SDKs are now generated from that schema, so they cannot drift apart. New in 1.0:
- Human-in-the-loop interrupts: an agent can pause a run, wait for a user decision, and resume
- Subagents: events carry a subagent ID, so the UI can show which worker did what
- Token usage: input, output, cached and reasoning tokens reported per run
- Multimodal tool results: images, audio, video and documents
- Metadata for passing custom data to the frontend
CopilotKit says the 1.0 spec will not change, and that it is backwards compatible in both directions: a 0.x agent works with a 1.0 client, and the reverse. The post lists framework support including LangChain, Mastra, OpenAI Agents SDK, Microsoft Agent Framework, Pydantic AI, Google ADK and Anthropic's Claude Managed Agents. CopilotKit also says Google, Microsoft, Amazon and Oracle have adopted the protocol. That is the vendor's own claim; we have not seen each company confirm it separately. CopilotKit's own v1.76.0 release moved onto the 1.0 packages on October 1.
Why a stable agent UI protocol matters for your business
Glue code is where agent projects quietly get expensive. Streaming partial text, showing tool progress, asking the user "approve this refund?" mid-run: every framework does it differently. Custom code for each means a rewrite every time you change frameworks or models.
A frozen spec is a portability guarantee. Build your frontend against AG-UI events and you can swap the backend, from LangChain to the OpenAI Agents SDK to Claude Managed Agents, without touching the UI. That is the same reason we build on open protocols everywhere else.
Interrupts are the feature that matters. The approval step is the most important control on any agent that acts. Having it in the protocol means it is a standard event, not a custom hack that breaks on upgrade.
Token usage in the stream is free cost tracking. You can log spend per run, per customer, from the event stream you already consume.
Key takeaways
- AG-UI 1.0 launched September 30 as a stable, open protocol connecting agents to app frontends
- 1.0 adds interrupts, subagents, token usage, multimodal tool results and metadata
- SDKs for TypeScript, Python and .NET are generated from one JSON Schema
- Backwards compatible with 0.x in both directions; CopilotKit says the spec will not change
- Build agent UIs on the protocol so you can swap frameworks and models without a frontend rewrite
Your agent's UI should outlive your agent's framework. We build agent frontends on open protocols, with approval steps and cost logging included, so you can change models without starting over. See how we build, or look at what we have shipped.
Sources: CopilotKit: Introducing AG-UI 1.0, CopilotKit v1.76.0 release on GitHub.
- #ag-ui
- #copilotkit
- #ai-agents
- #agent-protocols
- #frontend
Tommy Rush — Founder, Rush Commerce
Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More
Get The Rush Report weekly — one email, zero fluff.
Keep reading
GLM-5.3 built a Chrome exploit for $20. Patch in days.
Anthropic's red team found open-weight GLM-5.3 nearly matches Claude Mythos Preview at building exploits, for $20.40. Your patch window just shrank.
Read itAI coding tools fail WCAG: all 15 test sites missed Level A
AudioEye had five AI coding tools build 15 WCAG 2.2 AA websites. All 15 failed Level A, averaging 55 issues per page. Audit AI-built code before it ships.
Read it