The Collection

Five days on Grok Bot with a Cursor Pro+ account. Plugin catalog, browser login, connected. Freshdesk through the virtual browser became a fifteen-minute ticket check without a native connector. The letter’s frame: Grok Bot feels like unboxing a MacBook. OpenClaw feels like Linux. OpenClaw 2.0 narrows the gap with Quick Start and a browser app, but the split holds. Grok Bot is a managed agent computer. OpenClaw is a user-owned agent platform.

Programming moved up again. The interface is English. The thing being programmed is a Bot with a name, a role, and a description. An Agentic Engineer routes visual work to Claude Code, careful reading to Codex, simpler tasks to Grok Build. Personification is a cognitive boundary, not garnish. One hard line in the letter: every Bot shares the same computer, files, browser sessions, and logins. Separate Bots are organizational boundaries, not security boundaries. Update rebuilds while keeping durable state. Reset returns to the last sync and can drop unsynced work. No model picker. Useful as a digital chief of staff. Doubt that it authors most of your pull requests soon.

Separate bots are not security boundaries.

You hired a staff. They still share one machine.

Kotler keeps meeting business leaders who study “agentic AI” as if there is a final exam. Generative is the chat box. Agentic is Jarvis: multi-step, acting on its own. The fever, he writes, is bonkers. The Six Ds still map the curve. The useful crossing is Deception to Disruption, and the clear line is a user-friendly interface. Agentic AI is heading toward that border. It is not there. Current interfaces suck.

CEOs need the implications for the firm. That is not nine months becoming an AI mechanic. Friendly interfaces are in the pipeline. Grok has an early version. Everyone else is racing. Spending nine months on today’s agents for a three-month edge, if you even deploy it right, is the wrong trade. The 1994 analogy: understand what the internet meant for distribution and customers. Do not spend nine months learning HTML. The job of a CEO is still serving customers.

Understand the hire. Do not become the wrench.

The staff arrives when the interface stops feeling like a workshop.

Anthropic launched Claude as purpose-built add-ins inside Microsoft Excel and PowerPoint. Not an export hop. Not another tab. Inside Excel it reads multi-tab workbooks, explains formulas in plain English, debugs broken sheets, builds models and comp tables, cleans datasets, updates assumptions without snapping dependencies, and tracks changes with explanations. Inside PowerPoint it builds from templates, turns raw data into slides, rewrites structure, converts bullets to diagrams, keeps brand formatting, and can draft a full deck from a natural-language prompt.

Shared context across the hop is the quiet claim: Claude remembers the Excel work when you move into PowerPoint. Honest limits in the same letter: it only reads and writes files already open; it cannot open closed documents on its own; chat history resets between sessions; PowerPoint instructions stay separate from Excel instructions. Treat it as a brain layer on tools you already have, not a background hire that runs the office alone.

The loop was the tool. The hire sat inside it.

Finance, consulting, marketing, ops: the same rebuild that used to burn the afternoon.

Zvi’s capabilities letter opens on a weird week: Claude Fable 5.1 and GPT-6 Astra both arrive as “world’s most powerful,” both excellent. Fable 5.1: reduced cache reads ($1 to $0.25 per million tokens), zero data retention for eligible customers, classifiers less harsh than Fable 5 (biology false positives down about 85%, Claude Code cyber interventions about 60% fewer per session). Writing improved. Proactive. Early signs, large error bars: Sol to Astra may be the bigger jump than Fable 5 to 5.1. Zvi’s plan for non-trivial queries: dual wield.

Free Agents Weekly strip on Astra: built for operating a computer, not chatting. OpenAI launch comparison claims tops on seven rows. AutomationBench 41.4% versus 18.1% GPT-5.6 Sol and 31.4% Fable 5.1. Terminal-Bench 4.0 at 57.9% versus 37.3% and 55.8%. Terminal-Bench Science 0.1 at 64.6%. FrontierMath Tier 4 (v2) at 97.6%. ARC-AGI-3 at 99.9% versus 7.8% for Sol; no Fable 5.1 score reported there. Rollout: limited orgs day one, then Plus, Pro, Business, Enterprise, API, and AWS. Zvi’s early poll near his Twitter: Claude’s prior ~2:1 over Sol looks almost even now. Instinct’s free letter claims a personal agent valued at $2.5 billion the same day Linear hit that mark, $350 million raised, text-or-call, invite-only. The setup playbook sits unread behind the wall. We do not invent it.

Two candidates. One desk. Dual-wield until the hire earns the seat.

The model war is a hiring panel. Keep both on the roster while the dust settles.