I Analyzed 512,000 Lines of Leaked Code. It Shows What's Coming for Your AI Tools.

The biggest takeaway from the Claude code leak isn’t the leaked lines of software—it’s the outline of an “always on” agent platform called Conway, and how it could lock organizations into Anthropic’s ecosystem by turning workplace behavior into persistent, proprietary context. Conway is described as a standalone environment inside Claude with its own extension system, tool connectors, and event-based “wake up” triggers. The practical promise is that an agent can run overnight, monitor emails and Slack, pull context from internal docs, draft responses for approval, and prepare meeting materials—without waiting for a user prompt. The strategic implication is sharper: once an agent learns how a specific team works, switching providers becomes far more painful than migrating files or chat logs.

Conway’s interface is laid out as a dedicated sidebar page with three core areas: search, chat, and a “system” section that functions like an app store for agent capabilities. In that system area, users can install extensions packaged in a specific format (CNW.zip), connect external services, and configure automatic triggers via public web addresses that outside systems can ping to wake the agent. A key detail is that Conway’s extension ecosystem sits on top of Anthropic’s Model Context Protocol (MCP) but uses a proprietary packaging layer, making extensions effectively “Conway-only” rather than universally portable. That design mirrors the pattern behind Google Play Services: open foundations can earn credibility, while proprietary layers capture the economic gravity.

The leak also fits into a broader, quarter-by-quarter platform push. Over the past 90 days, Anthropic has rolled out Claude Code channels (messaging via Discord and Telegram), “Clawed Co-Work” for non-technical enterprise users, and a Cloud Marketplace procurement layer for partner apps. Then comes enforcement: third-party tools tied to Claude subscriptions are being restricted, with claims that pay-per-use rates could become dramatically higher if users rely on tools outside Anthropic’s own surfaces. Conway, positioned as the persistent agent layer, would sit on top of that harness—making the stack feel less like separate products and more like a single enterprise operating environment.

The transcript frames the trade-off that will define adoption. Flashy demos show agents that “just work,” but real value depends on speed, iteration, and the ability to supervise errors—because drafts and technical replies will sometimes be wrong. Still, even partial correctness can be enough if the agent is proactive and fast. That’s why the platform strategy matters: the money is not only in the model, but in the persistent memory of how people work—patterns learned over months of inference. Unlike data portability, behavioral context portability lacks clear legal or regulatory frameworks, and the “model of you” built from data plus compute may not be exportable in a way that preserves usefulness.

By late 2026, the transcript argues, the industry’s competition is shifting from model quality toward who owns the interface and the always-on memory layer behind it. Conway is presented as Anthropic’s “Active Directory” move for agents: it makes the rest of the stack sticky. The broader question for enterprises becomes whether to keep agent memory inside one vendor’s infrastructure or invest in a universal context layer that multiple models can access. The transcript ends with a warning that employee and career dynamics may change as employers increasingly encourage participation in proprietary agent platforms—turning workplace behavior into an asset that can compound for the company, even as individuals gain productivity from the system.