November 2026 Summaries

Functional copilots are enterprise AI agents designed for a specific business function and embedded directly in the software and workflows where users already work, enabling them to handle tasks such as contract reviews, approvals, and access requests while retaining human control over final decisions. Using a legal-contract example called Open Harvey, the post describes a five-step approach: researching workflows, defining accessible context and gated actions, specifying and building the application with CopilotKit and AG-UI hooks, securely deploying it, and adding governance measures such as approval gates, decision logs, authentication, audit trails, and evaluations. Rather than merely answering questions in chat, these copilots can read documents, apply policies, draft changes, route approvals, and preserve organizational knowledge as versioned, human-approved “skills.” The post also introduces CLiF, or Continual Learning from Full-Stack Trajectories, an open format that combines agent actions with users’ clicks, edits, and decisions to generate product insights, reusable skills, evaluation cases, and training data. CopilotKit provides an open-source repository and starter applications for building such copilots, emphasizing that the goal is to create a shared work environment in which agents support, record, and extend human judgment rather than replace it.

CopilotKit announced that Google Antigravity, Google’s Gemini-focused agent SDK and execution harness, is now compatible with the AG-UI Agent-User Interaction Protocol through the ag-ui-antigravity integration. The integration allows Antigravity agents to stream text, reasoning, tool activity, sub-agent work, file edits, and shell commands to user-facing applications such as web apps, Slack, and Teams, while supporting user messages and approval workflows in return. It works with CopilotKit features including tool rendering, shared frontend state, generative UI with A2UI, attachments, voice and headless interfaces, persistent threads, and learning from user actions and agent traces. Developers can set up an agent through a Python FastAPI server, restrict file access to designated workspaces, connect it to a CopilotKit runtime through an HTTP agent, and present it through components such as CopilotChat. The announcement also points to live demonstrations in CopilotKit Dojo and AG-UI Dojo, OpenBot’s use of Antigravity for AI coworkers, and related examples such as a CI-repairing agent tutorial and an open-source Repo Doctor application.

CopilotKit announced integrations that make Google Antigravity agents available in user-facing applications through the AG-UI protocol, allowing developers to stream agent reasoning, tool calls, subagent activity, and custom Python tool results into compatible frontends. The first-party ag-ui-antigravity integration supports conversation history, shared harness processes, and live demonstrations in the CopilotKit Dojo, while OpenBot, CopilotKit’s self-hosted AI coworker application, now supports Antigravity alongside other agent frameworks and records browser actions in audit logs. The accompanying Antigravity cookbook demonstrates Repo Doctor, an agentic application that diagnoses failing GitHub repository CI, modifies code, reruns tests to verify fixes, and keeps changes local rather than pushing them to GitHub. Repo Doctor includes live diffs, A2UI test-result cards, optional human approvals, activity tracking, downloadable reports, persistent conversation histories, and a Google Cloud deployment architecture using Cloud Run, Secret Manager, and Cloud Storage.

Repo Doctor is a tutorial application that combines Google’s Antigravity agent platform, its AG-UI streaming adapter, and CopilotKit to diagnose and repair failing GitHub Actions CI workflows through an interactive web interface. Powered by Gemini, the agent clones a repository into an isolated workspace, examines CI results, installs dependencies, reproduces failures locally, edits code, displays diffs and test-result cards, verifies that the originally failing check passes, and creates a local fix branch without pushing changes to GitHub. Antigravity supplies built-in file, shell, web-search, and custom Python tool capabilities, while AG-UI streams model output, tool activity, shared state, and approval interruptions to a Next.js frontend backed by FastAPI. Users can approve or deny potentially risky operations such as installs, edits, shell commands, and pushes, while CopilotKit renders tool-specific UI, A2UI test-report components, a live side panel, and downloadable run reports. CopilotKit Intelligence can persist conversations across devices and provide usage and reliability analytics, while the complete application can run locally or as a private single-instance Google Cloud Run service with secrets in Secret Manager and per-conversation workspaces preserved in Cloud Storage.

CopilotKit has introduced a native Vue SDK, @copilotkit/vue, bringing Vue applications feature parity with its React offering through AG-UI, an open protocol connecting frontend interfaces to AI agents. The SDK includes prebuilt chat interfaces, frontend tool calling, reactive shared context, human-in-the-loop approvals, generative UI components, agent and thread management, attachments, voice support, and debugging tools, while allowing the same underlying agent to work across Vue, React, React Native, and Angular. The initial implementation was created and contributed upstream by Abel Enekes of contract-management platform fynk, who ported React tests, developed Vue equivalents, and produced a parity guide; the contribution totaled more than 62,000 lines across 244 files before being merged in May 2026. Since release, CopilotKit and community contributors have added fixes, accessibility improvements, thread features, and documentation, while fynk has deployed the SDK in production and reported reducing its codebase by roughly 11,000 lines after adopting the open protocol.