The Future of Agentic Work: Karpathy on How AI Colleagues Follow Claude Code
What comes after Claude Code is the integration of AI into teams as "colleagues."
This is already beginning to move within the trend known as Agentic Work.
The way we use AI is about to move beyond the stage of competing over "which model is the smartest."
Which is better, ChatGPT or Claude? Which model won the latest benchmark? Which should be used for text generation, image generation, or programming?
Until now, the evolution of generative AI has been discussed primarily as a "competition of model capabilities."
However, following Andrej Karpathy's insights from May 2026 onwards, it becomes clear that the real point of contention lies elsewhere.
What is changing now is where AI is placed and the roles it is given.
Previously, AI was something where a human opened a website, wrote a question, and received an answer. From there, AI like Claude Code entered computers and terminals to manipulate files and code. In the next stage, AI will embed itself into Slack, codebases, internal data, business tools, and team conversations, working continuously within the organization.
AI is evolving from a "chatbot that answers questions" to an "agent that can be entrusted with work," and further toward a "colleague shared by the team."
This shift is symbolized by Karpathy joining Anthropic and the debut of Claude Tag.
Why Karpathy is a Special Figure in the AI Industry
Andrej Karpathy is not just a famous AI researcher. At Stanford University, he was involved in designing and teaching "CS231n," a course on computer vision and deep learning. Later, he conducted research as a founding member of OpenAI and led the AI department at Tesla, working on computer vision for Autopilot.
Furthermore, he has released educational materials like micrograd, nanoGPT, and "Neural Networks: Zero to Hero," conveying the inner workings of neural networks and LLMs in a way that can be understood through code. He possesses the rare ability to both understand advanced research and break it down for general developers.
In 2025, Karpathy popularized the term "vibe coding," becoming a symbol of a new development style where software is created by instructing AI in natural language. Since then, he has proposed the concept of "Agentic Engineering" to describe more full-scale software development.
On May 19, 2026, Karpathy joined Anthropic. He stated that the next few years at the LLM frontier would be a particularly important formative period, making it clear he was returning to the front lines of R&D. He was assigned to the pretraining team, which builds the fundamental capabilities of Claude.
Furthermore, Anthropic stated that Karpathy would launch a team to accelerate pretraining research itself using Claude. This isn't just career news. Karpathy, who had been focusing on AI education and personal projects, returned to the training grounds of cutting-edge models—and he chose Anthropic. He isn't just commenting on Claude from the outside; he is now on the side building its foundational power.
Why "Pretraining" is Crucial
When people think of generative AI performance, they often focus on new user features or improvements in how it answers. However, the capabilities of large language models are built in two major stages.
The first is pretraining. The model learns patterns about language and the world from vast amounts of text, code, images, and other data. This is where the foundation of what the model knows, its level of reasoning and abstraction, and the breadth of domains it can handle is formed.
The second is post-training. This involves adjusting how it follows human instructions, the safety of its answers, its reasoning processes, and its use of tools. This stage is heavily involved in the user experience and the model's "personality."
Karpathy joined the pretraining side. This means he is working on the foundation of next-generation Claude's knowledge, reasoning, learning ability, and world understanding, rather than just surface-level adjustments.
Even more important is the part about using Claude for pretraining research itself. AI researchers examine training data, write experimental code, analyze results, find causes of failure, and form new hypotheses. If Claude supports these research processes, AI will speed up AI research.
This isn't a simple story of "AI creating the next AI on its own." At this point, human researchers set goals, design experiments, and evaluate results. AI accelerates each of those steps. Still, if the research cycle shortens, a recursive loop is born: a stronger Claude supports research → research speed increases → an even stronger Claude is built → that Claude supports the next research.
Karpathy's move to Anthropic is significant because he has placed himself right in the middle of this loop of "accelerating AI research with AI."
AI is Ceasing to be a "Search Box"
Until now, the representative use of generative AI was Q&A. A human opens ChatGPT or Claude, enters a prompt, the AI returns an answer, and the human reads it. In this format, AI is essentially passive. It does nothing unless a human starts a conversation. Context must be explained every time. After receiving an answer, the human must move to another app to proceed.
For example, if you ask an AI for market research, it might write a report, but moving it to Notion, sharing it with the team, creating tasks, and assigning owners was a human job. The same applied to programming. Early AI coding support suggested code or wrote functions, but humans had to copy-paste, run it, and report errors back to the AI.
However, with Claude Code, this relationship changed. Claude can access the terminal, file system, codebase, and test environment to read, modify, run, and verify code itself. AI changed from "something that generates answers" to "something that works on a computer."
And in June 2026, Anthropic showed the next step: Claude Tag.
What is Claude Tag?
Claude Tag is a system that allows Claude to participate as a Slack team member. Administrators set which Slack channels, tools, data, and codebases Claude can access. After that, team members can tag @Claude on Slack to request work, just like calling a human colleague.
Claude breaks down the request into steps, uses authorized tools to work, and returns results to the Slack thread upon completion. It launched in beta for Claude Enterprise and Team users on June 23, 2026.
However, simply describing it as "using Claude from Slack" misses the true importance of Claude Tag. Services linking Slack and AI have existed for a while. Claude Tag has four distinct properties:
- Sharing one Claude as a team: In normal chat AI, everyone has separate conversations. A's Claude and B's Claude have different contexts. In Claude Tag, everyone in the same channel interacts with the same Claude. The whole team can see what Claude is doing, and other members can add instructions or take over previous conversations. AI work changes from a private chat to a shared team activity. Anthropic calls this "multiplayer."
- Continuous accumulation of channel context: Claude Tag accumulates information related to work from conversations in authorized channels. Project goals, past decisions, design judgments, team-specific terms, and assignments—the need to explain these from scratch decreases. Tacit knowledge often lives in Slack threads rather than formal docs. Claude Tag uses this scattered conversation as work context.
- Asynchronous work without human waiting: In traditional chat, users wait in front of the screen. With Claude Tag, you can assign a task and move on. Claude can work for hours or days, reporting progress to Slack. You can delegate different investigations to multiple Claudes in parallel. This changes the relationship from "Q&A" to "delegation."
- Acting on its own when necessary: Claude Tag has "ambient" behaviors, such as notifying users of changes in important metrics or following up on stalled tasks. AI becomes a presence that observes the environment and participates when needed, rather than just waiting to be called.
Why Karpathy Called it the "Third UI of LLMs"
Karpathy evaluated Claude Tag as the "third major redesign" of LLM UI/UX. In his framework:
- The first form is the Website: Accessing ChatGPT or Claude.ai to talk. Humans go to where the AI is.
- The second form is the App: Claude Code or coding agents. AI is installed on your computer to help with individual work.
- The third form is the Harness: AI becomes self-contained, persistent, asynchronous, and possesses organizational context. AI comes to where humans are already working (like Slack).
Karpathy describes Claude Tag as an organizational "harness." While making a basic demo is easy, making it work in an enterprise requires integrating tool links, execution environments, memory, security, permissions, auditing, and cost management into one system.
The "Harness" Becomes More Important Than the Model
A harness is the system that makes a model usable for actual work. No matter how smart a model is, it cannot do corporate work alone. It needs a system to determine what it can read, what tools it can use, what requires human approval, how it remembers the past, and how it handles failures. The harness gives the "brain" (the model) eyes, hands, memory, and a desk.
Claude Changes from a "Tool" to an "Agent"
This doesn't mean AI has become human. It means AI is approaching being a "subject of work." Traditional tools only work while a human operates them. With agentic AI, humans provide goals rather than operation commands. AI decomposes the goal, chooses tools, executes, and fixes problems. What returns to the human is not just a text answer, but a finished product: code, a pull request, or an analysis report.
Claude Opus 4.8 and "Long-term Working AI"
Just before Claude Tag, Anthropic released Claude Opus 4.8 on May 28, 2026. It featured "effort control," allowing users to adjust how much computation/reasoning Claude uses for a task, and "dynamic workflows" in Claude Code. Dynamic workflows allow Claude to create orchestration scripts to run dozens or hundreds of sub-agents in parallel, shortening weeks of work into days.
From Vibe Coding to "Agentic Engineering"
Karpathy noted that while "vibe coding" lowered the floor for software development, "Agentic Engineering" raises the ceiling for what elite developers can achieve. In this world, humans define what to build, verbalize requirements, set context, and review AI output. It is an engineering discipline of managing powerful but probabilistic AI without sacrificing quality.
Software 3.0: Context is the Program
In Software 1.0, humans wrote explicit code. In Software 2.0 (Neural Networks), humans provided data and objectives. In Software 3.0, LLMs are programmable via natural language. Context windows become the program. Development power will be determined by the ability to design what information goes into the context and how to verify the model's judgment.
From "Text Output" to "Interface Generation"
Karpathy also highlighted that AI output is moving from raw text to Markdown, then to HTML, and eventually to interactive interfaces. Instead of just explaining a policy, AI might generate a simulator where you can change tax rates to see the results. AI's role expands from "writing answers" to "generating environments for human understanding."
The Claude Fable 5 Incident: Capability and Governance
In June 2026, Anthropic announced Claude Fable 5, which Karpathy praised for its qualitative leap in handling long, complex tasks. However, it was suspended just three days later due to US government export controls regarding foreign national access. This event symbolizes that as AI moves from "giving information" to "exercising capability in the real world," it becomes a matter of national security.
AI Exponential and Recursive Improvement
Anthropic views this rapid progress as "AI Exponential." By May 2026, over 80% of code merged into Anthropic's own codebase was created by Claude. While humans still set the research themes and safety boundaries, AI is already accelerating the loop of creating the next AI.
Is Human Expertise Obsolete?
Anthropic's research suggests that domain knowledge is more important than ever. People who understand the business domain are better at catching AI mistakes and providing the right context. AI doesn't replace expertise; it amplifies it. The value of "converting specs to code" is dropping, but the value of "identifying the real problem" and "setting correct evaluation criteria" is rising.
The Power to Delegate
In the age of Agentic Engineering, the ability to delegate becomes a core technical skill. This means being clear about goals, constraints, completion conditions, and verification methods. It's similar to management, but you are managing a process executed by multiple AIs.
Conclusion: AI as a Persistent Working Subject
Karpathy's message in 2026 is that AI has become a new programmable layer. It is moving out of the chat box and into our organizations. The real competition is no longer just about who has the smartest model, but who can design the context, permissions, and organizational culture where AI can work safely and effectively. Claude Tag shows that this future is no longer a prediction—it is a product.
