For many years, AI integration into software was limited to a chat window tucked into the corner of the application interface. Users would type in a question, the AI model would return a piece of text, and then they had to manually translate that result into a real-world action. This approach was useful, but still passive.
According to CopilotKit, a Seattle-based startup co-founded by Atai Barkai and Uli Barkai, that model is no longer suitable for the modern generation of AI agents. Instead of simply chatting with users, AI needs to be integrated directly into applications, understand the context of the ongoing work, be able to perform actions, and display appropriate interfaces instead of just returning long texts.
In 2026, CopilotKit introduced a series of new technologies aimed at addressing the three biggest problems hindering AI agents from moving from the demo phase to real-world deployment: knowledge retrieval, testing reliability, and runtime persistence.
AG-UI: The missing piece in the AI Agent ecosystem
To understand CopilotKit's new tools, we first need to look at the protocol layer that the company is building.
Over the past few years, the AI agent ecosystem has gradually formed a relatively clear three-tiered architecture. At the bottom tier, the Model Context Protocol (MCP) helps agents connect to external tools, databases, and services. The next tier is A2A (Agent-to-Agent), responsible for coordinating and exchanging information between different agents.
However, a crucial gap remains: how does the agent interact with the user within the application?
Developed by CopilotKit, AG-UI focuses on the communication layer between the user, the application, and the AI agent. If MCP acts as the data connection system and A2A coordinates the agents, then AG-UI is the display and interaction layer that the user directly sees.
This protocol supports real-time feedback, creates dynamic interfaces, synchronizes state bidirectionally between the agent and the application, and enables the deployment of human-in-the-loop mechanisms, meaning the AI will pause and wait for user confirmation before performing crucial actions.
Currently, AG-UI has received support from many big names such as Google, Microsoft, Amazon, and Oracle. At the same time, this protocol has also been integrated into popular frameworks such as LangChain, Mastra, PydanticAI, and Agno.
Beyond theory, AG-UI now has official SDKs for various platforms and programming languages such as Kotlin, Go, Dart, Java, Rust, Ruby, and C++. Several projects supporting .NET, Nim, Flowise, and Langflow are also under development.
Notably, the Amazon Web Services (AWS) platform has integrated AG-UI into its FAST (Fullstack AgentCore Solution Template) and Bedrock AgentCore implementation examples, demonstrating that the protocol is gradually being viewed as a real-world infrastructure component rather than just an experimental idea.
If we liken MCP, A2A, and AG-UI to traditional web protocols, CopilotKit suggests that the relationship between them is similar to TCP, HTTP, and HTML. In this view, AG-UI acts like HTML – the display and interaction layer that end users directly use.
AIMock: When AI test suites might be "lying" to you.
One of the most notable products that CopilotKit will release in 2026 is AIMock.
According to CopilotKit, most current AI agent testing systems suffer from a major problem: they don't truly reflect real-world operating environments.
A request from a modern AI agent can go through a series of different components such as:
In practice, many development teams only mock one or two components, while still using the actual service. This results in inconsistent test results that are difficult to accurately reproduce across different environments.
AIMock was created to solve that problem.
This tool allows you to simulate the entire AI service chain through a single JSON configuration file. It supports up to 11 different LLM providers, including OpenAI, Claude, Gemini, Bedrock, Azure, Vertex AI, Ollama, and Cohere. Additionally, AIMock supports MCP, A2A, AG-UI, vector databases, and various APIs for searching or moderating content.
The three most outstanding features of AIMock include:
This allows development teams to detect many potential problems before the product is released into the real-world environment.
Interestingly, AG-UI itself now uses AIMock to test its protocol suite.
Pathfinder: A Knowledge Infrastructure Dedicated to AI Agents
While AG-UI addresses the interaction problem and AIMock handles testing, Pathfinder is built to solve a different challenge: how the AI agent accesses the right information it needs.
In demos, agents are usually provided with perfect data. But in a real-world business environment, information is scattered in many different places such as internal documents, source code, Notion, Slack, or Discord.
Pathfinder is a self-hosted MCP server capable of indexing all those data sources and transforming them into a knowledge base that agents can access.
The system can collect data from:
One notable point is that Pathfinder not only uses semantic search but also combines it with traditional keyword research.
According to CopilotKit, using only search vectors is often ineffective for queries containing error codes, specific APIs, or precise technical identifiers. Combining both methods helps the agent find more accurate information in technical environments.
Pathfinder also supports fully local deployment via Ollama or transformers.js, allowing businesses to build internal AI systems without sending data externally.
A single configuration file pathfinder.yamlis sufficient to set up the entire system. When the source code repository on GitHub is updated, Pathfinder can automatically re-index via webhooks without manual intervention.
CopilotKit is bridging the gap between demo and reality.
When viewed individually, AG-UI, AIMock, or Pathfinder might seem like technical tools serving specific needs. But when combined, they form a complete solution for AI agent systems in real-world environments.
Pathfinder solves the knowledge and context problem. AIMock ensures the system can be reliably tested before deployment. Meanwhile, AG-UI helps build the interactive experience between the user and the agent.
According to CopilotKit, these are the components often overlooked in demos but are the reason why many AI agent projects fail to scale up.
The company says its tools are currently seeing millions of installations per week and are being used by many Fortune 500 companies.
The AI agent market is entering a more mature phase, where discussions no longer revolve solely around which model is the smartest, but focus more on operational infrastructure.
Through AG-UI, AIMock, and Pathfinder, CopilotKit is attempting to build the missing foundational layers between AI models and real-world applications. Instead of creating a completely new proprietary runtime, the company chooses to provide tools that can work with a variety of frameworks, cloud services, and architectures.
That's also why CopilotKit is gradually becoming seen as one of the most noteworthy infrastructure projects in the current AI agent ecosystem.