This video argues that the essence of Anthropic's Claude Code leak was not a mere source code exposure or security vulnerability. Rather, it revealed the sweeping platform strategy Anthropic is pursuing through an always-on AI agent called "Conway." Conway can learn how you work to boost your productivity, while simultaneously creating a "behavioral lock-in" that binds you tightly to the Anthropic platform—and this, the video warns, will reshape the competitive landscape of the AI era.

1. The Conway Leak That Went Unnoticed 🕵️‍♀️

In 2026, Anthropic's Claude Code leak attracted enormous attention. Most coverage focused on the source code exposure or security vulnerabilities, but something far more significant was hidden inside: information about an always-on AI agent Anthropic was developing called "Conway." 😲 The details were cleverly buried within the Claude source code and represented an internal project with no mention in Anthropic's official roadmap whatsoever.

"The most important thing in the Claude Code leak wasn't the code. Last week I talked about how important it is to understand the harness around Claude Code. Today I want to talk about how Anthropic is building an always-on agent called 'Conway.' This was part of what leaked. It wasn't announced—it was buried in the Claude Code source when Anthropic accidentally published 500,000 lines of code to a public registry last week due to a packaging error."

Conway is not a simple chat window but a standalone agent environment. It has its own extension format, can be activated by external events, controls the browser, and connects to a wide range of tools. Viewed alongside the other products Anthropic has released over the past three months, Conway's emergence makes clear that Anthropic is pursuing a new platform strategy we did not anticipate.

2. What Does Conway Actually Look Like? 🤖

According to the leaked information, Conway operates as a standalone sidebar within the Claude interface—not a chat window, but a complete agent environment. The dedicated page connected to a Conway instance is divided into three core areas:

• Search 🔍
• Chat 💬
• System ⚙️

The System section is particularly interesting: it contains an Extensions area where users can install various add-ons—custom tools, interface panels, and capabilities that help the agent understand new types of information. Think of it as an app store for agent functionality. Someone creates and packages an extension, and Conway can install it.

Conway also has a Connectors and Tools section showing which services are connected. It includes a toggle that lets Claude and Chrome connect directly to a Conway instance, and it supports automatic triggers that allow external services to activate the agent. Users choose which services to grant this kind of access.

3. Imagining a Tuesday Morning with Conway 🗓️

What happens when Conway is woven into your daily life? Imagine waking up on a Tuesday morning to find that Conway has been running overnight and has already handled the following. 📧 It identified three emails critical to your work and drafted replies to the easy ones. It flagged your VP's email as important but left it untouched, knowing you'll want to handle it yourself.

Conway also scanned the Slack channels you monitor, found a question about authentication architecture in an engineering thread, pulled relevant context from a design document you reviewed last month, and drafted a reply that is waiting for your approval. 📝 It spotted a competitor mentioned in competitive intelligence, cross-referenced it with the research you had underway, and updated that information. It also noticed your board prep meeting at 10 a.m. and pulled the latest figures from the dashboard to help you prepare. Through all of this, you have not typed a single word. 🤯

Of course, roughly a third of what Conway did overnight might be wrong. An email draft might misread the tone; a Slack reply might be technically inaccurate. But thanks to Conway's speed and ability to iterate, even an imperfect agent can have a positive impact on your work.

"You haven't typed a word yet. And maybe a third of what Conway did overnight is wrong. The email draft misread the tone. The Slack reply was technically inaccurate. You'll figure that out very quickly. But if Conway is even partially right, the speed alone will still make it net positive."

This imagined portrait of Conway demonstrates the potential of AI agents, but it also reveals the gap between demo and reality. In polished demonstration videos everything works perfectly, but in practice these agents still need a great deal of hand-holding. That is likely exactly why Anthropic has not yet shipped Conway: the ability to accurately understand what a user prioritizes within an ocean of information is not yet perfect. 🤔

4. Anthropic's 90-Day Platform Strategy 🚀

Conway is not an isolated product—it is the culmination of a sweeping platform strategy Anthropic has been executing over the past 90 days. During that period Anthropic made the following significant moves:

• Launched Claude Code Channels: By enabling messages to Claude Code through Discord and Telegram and delivering task-completion notifications, Anthropic neutralized OpenClaw's core appeal. 💬
• Launched Claude Co-work: Targeting non-technical users, it aimed at the 95% of enterprise employees who are not engineers. Early adoption reportedly outpaced Claude Code's own rollout. 🧑‍💻
• Launched the Cloud Marketplace: A layer for enterprise buyers that lets them purchase Claude-powered partner apps—from GitLab, Harvey, Snowflake, and others—through Anthropic. Anthropic foregoes fees and focuses on capturing market share. 🛒
• Invested $100 million in the Claude Partner Network: Accenture committed 30,000 specialists to Claude training, with Deloitte, Cognizant, and Infosys joining as key partners—a full enterprise systems-integrator lock-in strategy. 🤝
• Blocked third-party tools: Anthropic blocked all third-party tools from Claude subscriptions, meaning anyone wishing to access Claude through a tool Anthropic did not build must pay 10–50× higher rates. 🔒

These five moves are not isolated product decisions; they form one unified platform strategy. Developer tooling (Claude Code), enterprise tooling (Co-work), an always-on agent (Conway), a distribution layer (Marketplace), and an enforcement mechanism (third-party tool block) all point in the same direction.

5. The Microsoft Parallel 📊

This strategy echoes Microsoft's growth trajectory in the 1990s. Microsoft began by selling the DOS operating system, then took control of the desktop with Windows, dominated the application layer with Office, and locked in enterprises through Active Directory and Exchange. Each step was an individual product, but together they gave Microsoft control over how enterprises compute. The whole process took roughly 15 years.

Anthropic is trying to run the same playbook as a speed run. 🚀 From model provider to developer tools to enterprise platform to agent operating system—all in 15 months. Conway plays the role of Active Directory in this strategy: the component that makes every other part of the stack sticky, because a persistent agent comes to understand an organization in ways nothing else can match.

6. MCP Open Standard vs. Proprietary Layer 🚧

One of the most telling details in the Conway leak is the extension format, which exposes a core tension at the heart of Anthropic's strategy. Anthropic published its own open standard called MCP (Model Context Protocol)—adopted by OpenAI, Google, and even the Linux Foundation—designed as a universal connector for linking AI tools to data sources. The premise was that any AI client could communicate with any data source through this open protocol.

But Conway uses MCP while building a proprietary layer on top of it. Conway's .CNW.zip extension format sits on top of MCP and includes custom interface panels, information handlers, and tools—but they only work inside the Conway environment. They are not portable tools that run anywhere; they are Conway-only tools. 🔌

This mirrors the Google Play Services pattern. Android is open-source software, but the Google Play Services layer—Maps, Payments, Push Notifications, the Play Store—that makes Android commercially viable is proprietary. Technically you can build an Android phone without Google, but nobody does, because the valuable features live in the proprietary layer.

Similarly, MCP is the open foundation, and Conway's extension ecosystem is the proprietary layer on top. Anthropic publishes an open standard to earn trust, then builds valuable tooling in a format that only runs inside its own environment, securing a commercial advantage. 💰

7. An App Store Choice for Developers 📲

Developers building tools for agents now face two paths:

1. Build standard MCP tools: Create portable tools that work across all MCP-compatible clients—Claude, ChatGPT, Gemini, and more. That's architecturally sound, but there is no distribution mechanism. No app store, no special discoverability. It's like building a website in 2008 while everyone else is downloading iPhone apps.
2. Build Conway extensions: They only work inside Conway, but Conway has a built-in extension directory—an app store. When Anthropic ships Conway to millions of Claude subscribers, your extension will be discoverable inside the environment where users are already working. You don't need to persuade anyone to install something; you just need to be in the app store.

This is precisely the choice Apple presented to mobile developers in 2008 and 2009: build for the open web, or build native apps for the iPhone. We know how that turned out. The open web may have been the better long-term architectural choice, but the App Store was where all the money went, and so that is where all the app developers went.

The upshot: if you trusted MCP and assumed your tools would be portable across every platform, you're technically correct. But Conway's extension model creates gravitational pull toward Anthropic, and other model vendors will follow. We are heading toward a world where Anthropic's tools only work in Conway, and OpenAI and Google are building equivalent silos of their own. 😥

8. Why Behavioral Lock-In Is Different ⛓️

Previous technology platform lock-in was primarily about data. Microsoft locked users in with files, Salesforce with customer records, Slack with communication history. That data is painful to migrate, but technically possible to move. Export tools exist, specialist consultants are available, and even if switching costs months and tens of thousands of dollars, it can ultimately be done.

When Conway ships, it will lock in something entirely different: an accumulated model of how you work. 🧠 Not your files, but the patterns the agent learned by watching how you use those files. Not your Slack messages, but the agent's understanding of which messages you reply to within five minutes and which you ignore for three days. Not your calendar, but the knowledge that you always reschedule your 2 p.m. Thursday slot and that meetings with your VP always run long.

There is no CSV export for that model. There is no migration consultant for how you think. So when you switch from Conway to another agent after six months, you don't just lose the agent—you lose six months of accumulated knowledge and experience that made the agent useful. You are back to square one with a "brilliant stranger" who knows nothing about you and must be taught everything from scratch.

"When Conway ships, it will lock in something different. It will lock in an accumulated model of how you work. Not your files—the patterns the agent learned by watching you use your files. Not your Slack messages—the understanding of which messages you reply to in five minutes and which you ignore for three days. Not your calendar—the knowledge that you always reschedule your 2 p.m. Thursday and that VP meetings always run long."

This is a new depth of lock-in that has never existed before.

9. Intelligence Portability Is Uncharted Territory 🤯

This new form of lock-in is different from data portability, for which we already have laws and frameworks. This is a question of "intelligence portability." The model the agent built of you is the product of your data, their computing, and six months of inference.

• Who owns it?
• Can you take it with you?
• And if so, in what format?

No legal or regulatory framework yet addresses these questions, because we have never faced them before. The speaker believes that the accumulated behavioral model of how you work should be portable, and that technical solutions and consensus are urgently needed to make that possible.

"This is not about data portability. We have laws and frameworks for that. This is about intelligence portability. The model the agent built of you is a product of your data, their computing, and six months of inference. Who owns it? Can you take it? And if so, in what format?"

Policy around behavioral context portability needs to be in place before Conway ships. Without it, we may be left defenseless against this new form of lock-in.

10. The Future AI Competitive Landscape: Who Will Own the Always-On Layer? 🏆

The first era of AI competition focused on the models themselves—who had the best foundation model, the best benchmarks, the longest context window. But as the gap between frontier models narrows, the primary competitive axis is shifting.

As we move from 2025 into 2026, the question becomes: "Who owns the interface where people actually do their work?" Claude Code, Cursor, OpenClaw, Windsurf—the "harness wars" defined this chapter of the competition.

Now we are moving one step further, from interfaces to the persistence and memory that power those experiences. 💾 The defining question for the rest of 2026 is: "Who owns the always-on layer?" Not agents that merely respond to prompts, but agents that run continuously, accumulate context, react to events, and act autonomously—agents that know you not because you told them something, but because they observed, learned, and remembered you. That is the persistence problem that will determine the next year.

Google, Anthropic, and OpenAI have all reached the same insight: the model is a loss leader, and what you truly want to own is the persistent agent layer. 💰 The layer that holds your memory, context, workflows, and integrations is the product that generates the revenue. Whoever owns that layer will achieve a degree of customer lock-in unprecedented in software history—not because their model is better, but because they have made the switching cost unimaginable.

11. Choosing an AI Agent Platform Carefully 💡

Choosing an AI agent platform will be harder to reverse than any software migration that came before.

• The temptation of convenience: Products like Conway will be polished, elegant, and arrive on day one with a full extension ecosystem. If Anthropic, Google, or OpenAI makes all of this effortless and it is incredibly good from the very first morning, you will never want to switch to anything else.

  "Anthropic and Google and OpenAI are going to make it easy. If you get an agent that handles everything, it pulls you in, it's easy to onboard, you're already a Claude user and now you can use Conway, and when you wake up it's incredibly good on day one, you're not going to switch."

• Will your agent become your company's brain?: Everything your agent learns about your organization, workflows, decisions, and institutional knowledge will live inside Anthropic—or whichever vendor you chose. Switch vendors, and you are leaving your "brain" behind. 🧠
• The value of ownership: If you build something like an "Open Brain"—a universal context layer—your memory belongs to you and is exposed through an open protocol accessible to any model. There is a setup cost, but you gain the benefits of true ownership.

Convenience vs. ownership. For many organizations, convenience will likely win. But the central challenge of 2026 will be the contest between companies and individuals who own their own permanent memory layer and those who lock themselves into an agent system and find themselves trapped. 🥊

12. Impact on Careers and Organizations 💼

These shifts will profoundly affect individual careers and how organizations operate.

• Individual careers: Advancement will increasingly depend on how well you leverage a persistent agent layer within your company—which requires a willingness to dive in and use it from day one. When building a team or starting a business, you will need to decide which agent interface to commit to.
• Ownership of behavioral evidence: The behavioral evidence of how we work should belong to us. We can allow an employer to learn from it and benefit from it while we are there, but we should be able to take it with us when we leave—our permanent talent fingerprint should not be something the company can copy and continue to exploit after we go.
• A new employer–employee dynamic: All of this introduces new dynamics into the relationship between companies and their employees. An employer can demonstrate that an employee is twice as productive when working with a particular agent, and that creates a powerful carrot and stick.

  "You can look at this employee and say: 'We know how good you are at this, this, and this. We know how to promote you. We know how to give you what you want in order to stay. But we also know that you are effective because of this agent. And you are going to want to stay. Because you can prove you are twice as effective with this agent.'"

• A new form of lock-in: In the second half of 2026, the lock-in effect on employees within a given employer will become very strong. That means choosing your employer carefully will matter more than it ever has before—far more than choosing between Windows and Mac. You are choosing whether you want to work with an agent that understands you, that you are comfortable with, and that makes you two or three times more productive.

Closing: What Persistent Agents Mean for Our Lives 💭

The Claude Code leak and the emergence of Conway offer important clues about how persistent agents will permeate our lives. Anthropic, OpenAI, and Google are all moving in this direction. We must each decide whether to chase convenience or to fight for ownership of our own "behavioral evidence" and "intelligence." This is not merely a technical decision—it raises profound questions about individual careers, the future of organizations, and human identity itself. The answers will begin to reveal themselves over the coming months. Choose carefully, and do not lose sight of what is truly at stake. 🙏