A packaging error pushed 512,000 lines of Claude Code source to a public registry. Buried inside was something Anthropic never announced: an always-on agent called Conway that reveals a platform lock-in strategy we haven't seen before.
Imagine you have a really smart assistant that lives inside your computer. Right now, AI assistants like Claude or ChatGPT only wake up when you talk to them. They answer, then go back to sleep. They forget everything.
But what if the assistant never slept? What if it watched how you work every day, learned what emails matter to you, knew which meetings always run late, and started doing things for you before you even asked?
That's what Conway is. Anthropic (the company that makes Claude) was secretly building it, and we only found out because they accidentally published their private code to a place where anyone could see it.
Here's the tricky part: after six months of learning how you work, Conway would know you so well that switching to a different assistant would feel like starting over with a stranger. You'd lose all that "training." And that's exactly what Anthropic wants — because if you can't leave, you're a customer forever.
It's like if your favorite game saved all your progress on one console, and the save file couldn't transfer to any other console. You'd never switch.
Last week, a packaging error at Anthropic pushed roughly 512,000 lines of Claude Code source to a public registry. Most coverage focused on takedown notices and security flaws. But buried in the code was Conway — an unreleased, always-on agent environment that doesn't appear on any public roadmap.
Conway operates as a standalone sidebar inside the Claude interface. Not a chat window — a full agentic environment with three core sections:
The vision: Conway runs 24/7, learns your patterns over months, and proactively handles email triage, Slack monitoring, meeting prep, and research — before you type a word. Even if a third of its work is wrong, the net value is positive because of speed.
Conway doesn't exist in isolation. In the last 90 days, Anthropic shipped five coordinated pieces:
Nate B Jones compares this to Microsoft's 1990s arc — DOS to Windows to Office to Active Directory — compressed from 15 years to roughly 15 months. Conway is the Active Directory play: the piece that makes everything else sticky.
Anthropic published the Model Context Protocol (MCP) as an open standard — adopted by OpenAI, Google, and hosted by the Linux Foundation. Any AI client can talk to any data source through MCP.
But Conway's CNW.zip extension format sits on top of MCP as a proprietary layer. Extensions include UI panels, custom handlers, and tools that only work inside Conway. This mirrors the Google Play Services pattern: Android is open source, but the commercially valuable layer (Maps, payments, Play Store) is proprietary. It also mirrors Apple's 2008 App Store vs. the open web — native apps with built-in distribution won.
Developers face the same choice: build a portable MCP tool with no distribution, or build a Conway extension that's discoverable inside an app store used by millions.
Previous platform lock-in was about stuff — Microsoft locked you in by files, Salesforce by customer records, Slack by communication history. Painful to migrate, but possible.
Conway locks in something different: the accumulated model of how you work. Not your emails, but the pattern of which emails you answer in five minutes vs. three days. Not your calendar, but the knowledge that your Thursday 2 PM always gets rescheduled. There is no export format for this. No migration consultant for behavioral context.
After six months, switching means losing all that compounding. You're back to a "brilliant stranger." This raises entirely new questions about intelligence portability that have no legal or regulatory framework.
Jones frames AI competition in three eras: Era 1 was models and benchmarks (GPT vs. Claude vs. Gemini). Era 2 (2025-2026) is interfaces and harness wars — which just climaxed with the OpenClaw ban and Peter Steinberger's defection to OpenAI. Era 3 (rest of 2026) is persistence and memory — who owns the always-on layer. All three labs are converging on the same play: model as loss leader, persistent agent layer as the money product.