September 2026 Summaries

AG-UI 1.0 introduces a stable open protocol for connecting AI agents to user-facing applications through a standardized bidirectional event stream, addressing the inconsistent custom streaming logic required by different agent frameworks. The release includes a JSON Schema defining event fields, behavioral rules for event order and error handling, and generated TypeScript, Python, and .NET SDKs, with input from Anthropic, Pydantic AI, and TanStack contributors. New capabilities include identifiable subagent streams, custom frontend metadata, multimodal tool results for files and media, human-in-the-loop interrupts and cancellation states, and token-usage reporting that includes subagent costs. Designed to work across frameworks such as LangChain, Mastra, Google ADK, Claude Managed Agents, and OpenAI Agents SDK, AG-UI supports interfaces including web, mobile, terminal, and messaging apps, while CopilotKit provides client integrations for several frontend frameworks and chat platforms. Version 1.0 is backwards compatible with 0.x implementations, although SDK upgrades require limited naming, validation, metadata, and typing adjustments.

AG-UI (Agent-User Interaction Protocol) is an open, framework-agnostic standard for streaming AI agent activity to user-facing applications, addressing the difficulty of displaying real-time reasoning, tool calls, state updates, and parallel subagent work across different frameworks and interfaces. Reaching a stable 1.0 specification, it is supported by major companies and integrations including Google, Microsoft, Amazon, Oracle, Anthropic, LangChain, Mastra, and Pydantic AI, with official TypeScript, Python, and .NET SDKs. The protocol connects an application, client SDK, agent endpoint and bridge through transport-independent JSON events, typically sent using Server-Sent Events, while complementing MCP for tools and data and A2A for agent-to-agent communication. Its 31 event types cover runs and steps, streamed text, tool calls and multimodal results, reasoning, synchronized state, live activity, subagents, and custom or raw events, allowing interfaces to react progressively during a run rather than wait for a final response. AG-UI 1.0 adds features such as human-in-the-loop interruptions, token usage reporting, metadata, capabilities declarations, encrypted reasoning support, and backward compatibility that lets older clients ignore unfamiliar events. The guide illustrates implementation through agent endpoints and SDK clients, demonstrates a flight-booking workflow with live UI updates and approval steps, and points developers toward starter tooling, documentation, integrations, an event inspector, and CopilotKit-based clients for web, mobile, and messaging platforms.

AG-UI, or Agent-User Interaction Protocol, is presented as a standard for connecting AI agents with user-facing applications through structured events that communicate agent activity, shared state, tool calls, user decisions, and errors. Using an appointment-rescheduling example, the discussion identifies product decisions around what context an agent receives, how progress and failures are displayed, how users provide input, when approvals are required, and which frontend and backend actions agents may request while applications retain authorization and business-rule enforcement. A common protocol can help teams reuse interface patterns and support multiple agent frameworks, although integrations still require testing. The piece also argues that teams should evaluate complete user-agent trajectories rather than treating successful tool calls as sufficient, using records of user context, approvals, agent requests, and saved outcomes to detect failures such as booking an unapproved time. It recommends starting with a single existing workflow, defining clear acceptance criteria and recovery behavior, then using observed user interactions to improve the agent experience.

Microsoft has contributed a first-class C#/.NET SDK for the AG-UI Agent-User Interaction Protocol, bringing version 1.0 packages to NuGet and consolidating protocol support previously embedded in Microsoft Agent Framework. AG-UI standardizes the event stream between AI agent backends and user-facing clients, covering lifecycle activities such as streamed text, tool calls, state changes, subagents, interrupts, and run completion, enabling clients to work independently of the backend language or framework. The SDK includes server tooling to expose any .NET IChatClient-based agent as an AG-UI endpoint and client tooling that lets .NET applications consume endpoints written in C#, TypeScript, or Python, with Server-Sent Events as the default transport and optional protobuf support. Microsoft Agent Framework now relies on these shared packages while retaining ASP.NET Core integrations, with minor API renames but unchanged wire compatibility for existing frontends. The release enables .NET services, internal APIs, and business applications to expose agents across web, terminal, mobile, Slack, and Teams interfaces without requiring a specific agent framework.

CopilotKit describes using Jev, TypeSafe’s non-generative “System One” decision model, to create generative user interfaces from an organization’s existing, approved component catalog rather than having language models generate new markup. Jev accepts application state and structured questions, then returns probabilistic typed decisions through Choice, Noul, and Score primitives that can select components, evaluate conditions, and determine display priorities in parallel. In the proposed Generative UI Decision Engine workflow, frontend components are registered with CopilotKit and sent through the AG-UI protocol, Jev selects among them, the choices are assembled into an A2UI JSON layout, and a text model subsequently fills content slots. Examples involving insurance claims, banking dashboards, and internal IT portals show how the approach can produce context-specific interfaces while maintaining design-system, accessibility, and compliance constraints. The post also frames the process as catalog-based classification, enabling confidence thresholds, default fallbacks, probability logging, and explainable UI choices, and suggests beginning with a single text-heavy agent screen and a small set of registered components.

CopilotKit’s September 10, 2026 session demonstrates how developers can build an AI agent once and deploy it across React applications, Slack, and Microsoft Teams using CopilotKit and the AG-UI protocol. Rather than maintaining separate implementations for each platform, AG-UI streams typed events such as messages, tool calls, state updates, and approval requests, while each surface renders those events through its native interface, including React components, Slack Block Kit, and Teams Adaptive Cards. The presentation emphasizes that shared state and consistent human-in-the-loop decisions are essential to preventing cross-platform behavior from diverging, allowing approvals or edits made in one surface to be reflected everywhere. It also covers streaming, tool lifecycles, generative UI, channel context, and the limited surface-specific code required at the rendering layer, arguing that protocol-based agent architectures are easier to extend than agents tightly coupled to a single UI. The setup, documentation, AG-UI protocol, and CopilotKit source code are available as open-source resources.

CopilotKit has released an upgraded Inspector, a development-only debugging tool for AI agents that works across React, Vue, and Angular and automatically activates in CopilotKit projects. Built around the AG-UI protocol, it provides real-time visibility into agent events, tool calls, context, failures, and lifecycle activity, while linking errors to their underlying events and allowing event streams to be searched, filtered, or exported. New features include an isolated Playground for testing agents without affecting production chat histories, conversation forking to reproduce user issues, bidirectional navigation between stored threads and live app chats, configurable Inspector placement, and setup prompts for optional capabilities. When connected to CopilotKit Intelligence, the platform adds persistent rich threads, usage analytics, user memories, and Automatic Learning, which analyzes conversation patterns and proposes reviewable SKILL.md files for agent improvements. Intelligence can be used through a hosted developer tier or self-hosted deployment, while the Inspector itself remains available independently and never loads in production or server-rendered environments.

CopilotKit has introduced WebMCP support, allowing tools registered through its useFrontendTool API to be discovered and invoked by compatible browser agents such as ChatGPT Atlas, Comet, and Dia, as well as by an application’s own embedded copilot. Developers can opt individual tools into WebMCP with minimal configuration, automatic component-lifecycle registration, safety annotations, and support across frameworks including React, Vue, Angular, and React Native, with fallbacks for unsupported or server-rendered environments. The release also positions CopilotKit as a platform for building in-app agents that use these tools while supporting generative interfaces, persistent threads, shared application state, human approvals, and multiple agent backends. The announcement encourages developers to extend existing CopilotKit implementations, expose an appropriate read-only demonstration tool when needed, and verify browser discovery and invocation, while noting that the feature is open source and requires limited setup.

CopilotKit’s Slack Agents Night showcased how its Channels SDK can bring AI agents into workplace messaging platforms, with live demonstrations of the internal Kite agent and the open-source OpenTag Slack agent. Kite created a pull request from a Slack conversation, requested approval before publishing it, and used native Slack Block Kit components for status updates, charts, cards, media, and approval forms. The SDK supports more than 20 agent frameworks through AG-UI, managing authentication and middleware while allowing developers to write JSX components once for rendering across Slack, Microsoft Teams, and other supported surfaces. OpenTag, which has surpassed 1,200 GitHub stars, is presented as a production-ready Slack agent template that can connect to a custom backend without requiring changes to its channel layer. The post also highlights related CopilotKit updates, including OpenBot’s expanded enterprise and automation capabilities, full AG-UI compatibility for Mastra agents, and a React Native SDK for embedding agents in iOS and Android applications.

OpenBot’s v0.0.5 release follows the project’s surpassing 3,500 GitHub stars and expands its open-source, self-hostable platform for deploying AI agents as workplace collaborators. New capabilities include first-party Google Drive and Notion connectors with user-specific permissions, bot-to-bot delegation governed by administrators, scheduled routines that operate under their creators’ permissions, and automatic routing of untagged messages to the most suitable bot. The release also adds human browser takeover, context-sensitive tool selection through reusable skills, enterprise SSO and Kubernetes deployment support, and isolated bot environments that suspend when idle. Security enhancements include rule testing against historical activity, network and credential protections, and audit logging, while usability updates cover channel management, conversation recovery, support for newer reasoning models, and interface refinements. The post also promotes local setup resources and highlights related CopilotKit developments involving Slack and Teams agents, Mastra compatibility, and React Native support.