Most important take away

Buried in Anthropic’s accidental 512,000-line Claude code leak is “Conway,” an unannounced always-on agent that runs persistently, wakes on external triggers, and learns how you work over time. The real lock-in won’t be your files or messages but the accumulated behavioral model of you that the agent builds — a new kind of switching cost with no export format, no migration tool, and no legal framework yet.

Summary

What Conway actually is

• A standalone agent environment (not a chat window) hidden inside the leaked Claude code source, never announced on Anthropic’s roadmap.
• Opens as a dedicated sidebar with three areas: Search, Chat, and System.
• The System area includes: an extension store (C&W.zip format) for installing custom tools, UI panels, and data handlers; a connectors/tools section (including a toggle letting Claude + Chrome connect directly); and automatic triggers — public webhook URLs that let external services wake the agent.
• Designed to run overnight, monitor email/Slack/calendar/competitive intel, draft responses, and prep you for meetings before you type a word. Expect roughly a third of its output to be wrong — value comes from speed and iteration, not first-try accuracy.

The 90-day platform strategy Conway completes

Anthropic has shipped five coordinated moves this quarter that only make sense as one strategy:

1. Claude Code Channels — notifications via Discord/Telegram, neutralizing OpenClaw’s appeal inside Anthropic’s own surface.
2. Claude Co-Work — non-technical enterprise tool targeting the 95% of employees who aren’t engineers; reportedly outpaced Claude Code’s early adoption.
3. Claude Marketplace — enterprise procurement layer where partner apps (GitLab, Harvey, Snowflake) are bought through Anthropic against existing spend commitments. No commission yet — buying distribution.
4. $100M Claude Partner Network — Accenture training 30,000 pros; Deloitte, Cognizant, Infosys as anchor SIs. Classic SI lock-in at enterprise scale.
5. Blocking third-party tools from Claude subscriptions — OpenClaw first, everything else in coming weeks. Pay-per-use rates run 10-50x higher than subscription rates.

Nate frames this as Anthropic “speed-running” Microsoft’s 15-year arc (DOS → Windows → Office → Active Directory) in 15 months: model provider → dev tool → enterprise platform → agent operating system. Conway is the Active Directory play — the piece that makes everything else sticky.

The MCP vs. C&W.zip tension (the key strategic insight)

• MCP (Model Context Protocol) is Anthropic’s own open standard, adopted by OpenAI, Google, and the Linux Foundation.
• Conway uses MCP but layers a proprietary C&W.zip extension format on top. Extensions only run inside Conway.
• This is the Google Play Services pattern: open kernel, proprietary commercial layer where the value lives.
• Developer choice now mirrors 2008’s web-vs-iPhone-App-Store decision:
  • Build standard MCP tools: portable, but no distribution, no store, no featured placement.
  • Build Conway extensions: locked to Conway, but in-store discoverability to millions of Claude subscribers from day one.
• Known outcome: the proprietary store wins the money even if the open standard wins the architecture debate.

The OpenClaw pattern — Anthropic’s playbook

1. Copy the popular community feature into the closed harness.
2. Subsidize the first-party version inside the subscription.
3. Make the third-party version expensive or impossible.
4. Ship a proprietary format that pulls the ecosystem onto your surface. Peter Steinberger’s move to OpenAI on Feb 14 and the subsequent enforcement wave against OpenClaw is treated as evidence this is already underway. Conway is at step 1; steps 2-4 were visible in the leak.

A new kind of lock-in: intelligence portability

Previous lock-in was about stuff (files, records, messages). Painful but migratable in months. Conway locks in the accumulated behavioral model — not your calendar but the fact it knows you always reschedule 2pm Thursdays; not your Slack but which messages you answer in 5 minutes vs. 3 days. There is no CSV of “how this person thinks,” no migration consultant for behavioral context. Switching means going back to “a brilliant stranger” and losing six months of compounding. Nate’s core policy argument: portability policies for behavioral context must ship BEFORE Conway does, not after. Today there are no legal, regulatory, or even opinion frameworks for who owns the model-of-you.

Three eras of AI competition

• Era 1 (2023-24): Best foundation model. Margins have now compressed.
• Era 2 (2025-early 2026): Who owns the interface/harness (Claude Code, Cursor, OpenClaw, Windsurf). Climaxed with the OpenClaw ban.
• Era 3 (rest of 2026): Who owns the always-on persistence/memory layer. All three labs (Anthropic, OpenAI, Google) have converged on treating the model as a loss leader and the persistent agent as the money product.

Actionable insights

For builders/developers:

• Assume Conway-class proprietary extension formats are coming from all three major labs. Decide deliberately whether to ship via open MCP (portable, hard to distribute) or proprietary stores (locked in, instant reach).
• If you’re building on MCP expecting portability to matter commercially, understand the gravitational pull toward vendor-proprietary layers — other labs will copy this.
• If you’re building agent memory/context infrastructure, there’s a real window right now to offer a universal, portable context layer (Nate plugs OpenBrain as one approach). Conway’s launch will test whether anyone will pay the setup cost to avoid lock-in, or whether convenience wins.

For enterprise architects:

• Ask explicitly whether your agent’s accumulated organizational memory will live inside a single provider. The convenient answer (yes) means your institutional knowledge, workflows, and decisions become non-exportable.
• Negotiate behavioral-context portability into vendor contracts now, before persistent agents ship broadly.
• Expect SI lock-in (Accenture/Deloitte/Cognizant/Infosys trained on Claude) to make switching harder regardless of your technical stance.

For employers and HR/leadership:

• Conway-class agents empower employers more than employees in the relationship: the company can see exactly how effective each worker is, what makes them effective, and knows the agent is now part of that effectiveness.
• Have an intentional conversation now about how to allocate the value created when an employee’s behavior trains the company’s context layer. Options: compensate for it, let employees export a behavioral fingerprint, or (most likely, per Nate) default to “everything you do at work belongs to the company.”

Career advice (explicit and implicit)

• “Pick your fighter carefully.” Choosing an employer in the second half of 2026 will increasingly mean choosing which persistent agent ecosystem you want to work inside — Claude, ChatGPT, or Gemini. This is a bigger decision than Windows vs. Mac ever was.

• Promotion will become a function of how well you use your company’s persistent agent layer. That requires willingness to jump in from day one.

• Building a team will require leading across a persistent context layer your company chose — your team’s effectiveness is partly the agent’s effectiveness.

• Starting a business means picking which agentic interface you build on (or building your own).

• Lock-in for employees inside employers will get “very strong” in H2 2026 — the agent you’re productive with belongs to the company, and leaving means losing compounding productivity, not just a job.

• Nate’s personal recommendation: own your own persistent behavioral audit/memory layer if possible, so your talent imprint travels with you. He acknowledges that for many people this won’t be an option because employers will hold the cards.

• Bigger picture: don’t assume agents will cause mass firings — managing them still requires many smart humans. The real shift is in leverage and switching costs within the employment relationship.

• Current consumer/professional sorting Nate is already observing:

  • Broad consumers gravitating to ChatGPT Free/Go.
  • Active professionals gravitating to Claude/Anthropic plans.
  • This is starting to leak into enterprise contracts — the choice of persistent-agent vendor is becoming a career-relevant bet.

Chapter Summaries

1. The leak that buried Conway — Amid 512k leaked lines of Claude code, the biggest untold story is Conway, an unannounced always-on agent with its own environment, extensions, triggers, and tool connectors.

2. What Conway looks like inside the Claude interface — A standalone sidebar with Search, Chat, and System sections; the System section contains an extension store (C&W.zip), connectors/tools, and webhook-style automatic triggers.

3. A Tuesday morning with Conway — Concrete walkthrough of the overnight agent: flagged VP emails, drafted Slack replies, board-meeting prep — plus the reality that ~1/3 of outputs will be wrong, so value is speed-driven.

4. Conway as the capstone of a 90-day strategy — Five shipped moves (Channels, Co-Work, Marketplace, $100M partner network, third-party tool ban) form one coordinated platform play, not five products.

5. The Microsoft speed-run analogy — Anthropic is attempting the DOS → Windows → Office → Active Directory arc in 15 months; Conway is the Active Directory equivalent.

6. MCP vs. C&W.zip — the open-standard head fake — Anthropic published MCP as open, but Conway’s proprietary extension format sits on top, replicating the Google Play Services pattern and forcing a 2008-style developer choice.

7. The OpenClaw pattern as Anthropic’s playbook — Four-step pattern: clone, subsidize, lock out, ship proprietary format. Steinberger’s OpenAI defection and the OpenClaw ban show the pattern is active.

8. Intelligence portability — a new lock-in layer — Conway locks in the behavioral model of you, not files. No export format exists, no consultant can migrate it, and switching costs six months of compounding. Portability policy needs to ship before Conway does.

9. Three eras of AI competition — Models (2023-24) → Interfaces/harnesses (2025) → Persistent memory/agent layer (rest of 2026). All three labs are treating the model as a loss leader and the persistent agent as the product.

10. Enterprise and builder calculus — The decision: accept provider convenience and lose behavioral portability, or invest in a universal context layer. Nate predicts convenience will win for most companies.

11. Career and employer implications — Persistent agents empower enterprises in the employee relationship. Picking an employer now means picking an agent ecosystem; promotion, team-building, and entrepreneurship all get filtered through that choice.

12. Closing charge — Look past the OpenClaw drama and understand why Anthropic is building toward the persistent agent layer, because OpenAI and Google will ship their versions within months and the behavioral-portability conversation needs to happen now.