Anthropic Surface Expansion

Canonical wiki thread for tracking Anthropic’s coordinated push to own the persistent agent layer — its bet that the model is a loss leader and the hosted always-on agent is the money product. As of April 2026 this is one of the wiki’s load-bearing structural threads. Spawned from Nate B Jones’s Conway leak analysis which made the previously-distributed pattern explicit.

The Three Eras frame (Nate B Jones)

This thread sits inside a broader frame for how frontier-AI competition has evolved:

EraYearsAxis of competitionStatus
Era 1 — Models2023–24Foundation models, benchmarks, training runs, context windowsMargins compressed; no longer the primary axis
Era 2 — Interfaces2024–25Harness wars: Claude Code vs Cursor vs OpenClaw vs WindsurfJust climaxed with the OpenClaw ban + Steinberger’s defection to OpenAI
Era 3 — Persistence + Memory2026+Who owns the always-on layer that accumulates context, wakes on events, acts autonomouslyThe current axis. All three frontier labs converged on the same insight

“They’re building the model as a loss leader and they want to own the persistent agent layer. That’s the money product. Whoever owns that layer has customer lockin like we have never seen before — not because the model is better, but because the switching cost is unthinkable.”Nate B Jones

This analysis page tracks Anthropic’s specific moves in Era 3. For the cross-lab comparison (Anthropic vs OpenAI vs Google), see persistent-agent-layer-wars.

The 90-day platform play

Nate B Jones’s framing: Anthropic’s last quarter is not five separate product decisions but a single coordinated platform strategy speedrunning Microsoft’s 15-year DOS→Windows→Office→Active Directory arc in 15 months. The seven moves:

  1. Claude Code channels — Discord/Telegram messaging surface; task-completion notifications. Neutralized OpenClaw’s appeal inside Anthropic’s own product.
  2. Claude Co-work — non-technical-user surface targeting the 95% of enterprise employees who aren’t engineers. Adoption reportedly outpaced Claude Code at the same stage.
  3. Claude Marketplace — enterprise procurement layer (GitLab, Harvey, Snowflake purchasable through Anthropic; counts against existing spend commitments; no commission yet — buying distribution market share).
  4. $100M Claude Partner Network — Accenture training 30,000 professionals on Claude; Deloitte/Cognizant/Infosys as anchor SI partners. “Complete system integrator lockin at enterprise scale.”
  5. Third-party tool ban — January quiet block, February ToS revision, recent enforcement against OpenClaw first; pay-per-use rates 10–50× the subscription cover; rolling out to “everything else in coming weeks.”
  6. Managed Agents (April 2026) — first hosted agent runtime. The onramp tier: Claude.ai users build their first agent without infra work. $0.08/hr per active session + tokens. Three private-preview features (Outcomes, Multi-Agent Orchestration, Persistent Memory) define whether it becomes a real builder tier.
  7. Conway (leaked, unannounced) — the always-on capstone surfaced from the Claude Code leak. Standalone sidebar UI, proprietary .cnw.zip extension format on top of MCP, automatic webhook triggers, browser control, persistent memory across sessions. “The Active Directory play — the piece that makes everything else in the stack sticky because the persistent agent knows your organization in a way nothing else does.”

“Anthropic is speedrunning Microsoft. Model provider to developer tool to enterprise platform to agent operating system in 15 months. Microsoft took 15 years.”

The four user-facing surfaces

After the 90-day arc, Anthropic now has four primary user-facing surfaces, each targeting a distinct user:

SurfaceTarget userStatusKey wiki page
Claude.aiEnd users (chat, work tasks)GAclaude
Claude CodeDevelopers (CLI agent harness)GA, mature; /loop, Ultra Plan, skills ecosystemclaude-code
Managed AgentsClaude.ai users building their first agentGA, April 2026managed-agents
ConwayKnowledge workers needing always-on org contextLeaked, unannouncedconway

The two bottom rows are the wiki’s future-facing watch. Watch list:

  • Conway naming on a real Anthropic surface (docs, roadmap, launch post) confirms the leak
  • .cnw.zip published as an extension SDK is the “build for the App Store” inflection for tool developers
  • Automatic triggers shipping in Managed Agents would close the cron gap and reduce reliance on trigger.dev for always-on Anthropic-native workflows
  • Browser control as a first-class Anthropic feature would compete with Stagehand / OpenClaw / Computer Use
  • Any portability commitment before Conway launches is the rare counterexample to the four-step playbook

The four-step playbook

The OpenClaw / Steinberger / ban timeline exposes a repeatable pattern for how Anthropic absorbs community-built primitives — the wiki tracks this as the four-step playbook:

  1. Copy — build the first-party version of what the community built
  2. Subsidize — make the first-party version free or cheap inside the subscription
  3. Block — make the third-party version expensive (10–50× pay-per-use) or technically impossible
  4. Lock the format — ship a proprietary extension format (.cnw.zip) that ensures the ecosystem builds for the platform’s surface, not the open one

The Steinberger / OpenClaw evidence:

  • Feb 14, 2026 — Peter Steinberger (OpenClaw creator) joins OpenAI; Sam Altman publicly tags him to “drive the next generation of personal agents”
  • Within weeks — OpenClaw moves to a foundation with OpenAI backing
  • January 2026 (retroactively visible) — Anthropic quietly blocks third-party tools from using subscription login credentials
  • February 2026 — Anthropic ToS revision codifies the block
  • April 2026 — enforcement against OpenClaw first; rolling out to “everything else in coming weeks”

The OpenClaw → Claude Code channels → Managed AgentsConway arc is the canonical playbook target. Watch for OpenAI and Google running the same play in persistent-agent-layer-wars.

What the proprietary .cnw.zip layer means for MCP

The structural detail Nate B Jones keeps returning to: Conway extensions sit on top of MCP but are not portable MCP tools — they only work inside Conway. This is the MCP-as-open-foundation + .cnw.zip-as-proprietary-distribution pattern. The historical parallels:

Open layerProprietary layer that captured the value
Linux / AOSPGoogle Play Services (Maps, Payments, Push, Play Store)
Web standardsiOS App Store
MCP.cnw.zip extensions inside Conway

Anthropic gets the credibility of publishing an open standard and the commercial advantage of building the valuable tooling in a format that runs only in their environment. The wiki tracks this as the proprietary-layer-on-top risk — a flag for any future “open” agent standard that the value may shift to a proprietary store on top.

What this enables: behavioral lock-in

The conceptual payload of the surface-expansion thread is behavioral-lock-in: a new vendor lock-in surface that’s structurally different from anything before. Conway-class always-on agents accumulate the model of how you work — and there’s no export format, no migration consultant, no legal framework for portability. Every previous lock-in was about stuff (files, records, comms history); this is about the accumulated inference + observation a vendor’s compute did over months. See behavioral-lock-in for the full concept page; OpenBrain for the structural counter-pattern.

How the wiki tracks this thread

  • This page is the canonical entry point. Update on every new Anthropic surface, every Conway data point, every new playbook step.
  • anthropic keeps a 1-paragraph stub linking here, plus the product list. Don’t duplicate the 7-step play on the org page.
  • behavioral-lock-in holds the conceptual frame that this thread anchors.
  • persistent-agent-layer-wars holds the cross-lab comparison (Anthropic vs OpenAI vs Google).
  • Watch flags in wiki/hot.md — any Conway naming, .cnw.zip SDK, native triggers in Managed Agents, OpenAI/Google equivalent launches.
  • CLAUDE.md scope explicitly includes the persistent-agent-layer competition and lock-in/portability dimension.

See Also