Silicon Valley Goes “Claw”: Anthropic, Microsoft, Meta, Notion Rush to Build Their Own AI Agents

“Claw” denotes AI agents that can click, manipulate apps, and automate workflows, effectively acting as a pair of hands. In the past two weeks, major AI firms released their own Claw‑type products.

Meta – Manus Agent

Meta upgraded the Manus Agent and embedded it directly into Telegram. The agent emphasizes long‑term memory, aiming to retain a user’s style, tone, and fine‑grained preferences. Example: sending “make a video as usual” triggers the agent to retrieve historical assets, coordinate with Gmail and Notion, and complete the full video‑creation pipeline (script, generation, delivery).

Anthropic – Claude Cowork Updates

Anthropic accelerated releases within 48 hours, first publishing remote‑control code for phones and then unveiling an automated timed‑task system for Claude Cowork.

I expect Claude Cowork to be comparable to OpenClaw in 2‑4 weeks.

Microsoft – Copilot Tasks

Microsoft announced Microsoft Copilot Tasks , which can autonomously generate a weekly work plan from a user’s calendar, read Outlook emails to extract key details, automatically schedule meetings in Google Calendar, and draft PowerPoint outlines. It also supports cron‑style recurring tasks, e.g., “summarize today’s team progress at 5 pm and send the weekly report.”

Notion – Custom Agents (v3.3)

Notion released Custom Agents , a 24/7 autonomous system that can be triggered by conditions such as a Slack @‑mention. Agents audit every click and log via the MCP protocol across Slack, Figma, HubSpot, and Notion Mail. Early testers created over 21 000 agents, with 2 800 agents rotating inside Notion.

Perplexity – Perplexity Computer

Perplexity launched Perplexity Computer , an “all‑in‑one” AI that can research a topic, write code, and deploy it automatically, aiming to replace the traditional chatbot model with a full execution pipeline.

Why the Shift Now?

1. Model capability crossing the trust threshold. Earlier models excelled at art and poetry but struggled with multi‑step, high‑logic tasks. Newer models such as o1, which incorporate reinforcement learning and chain‑of‑thought reasoning, show a qualitative leap in logical consistency, allowing safe delegation of system permissions to agents.

2. Industry consensus drift. Scaling‑law gains are diminishing; merely increasing parameters yields little user‑perceived benefit. Companies now prioritize “execution‑centered” AI that can book flights, edit code, or manage files, turning AI from a knowledge hub into a productivity engine.

3. Business‑model evolution. Traditional AI chatbots monetize by selling tokens for generated text. In the agent era, revenue comes from selling concrete labor hours—AI that performs real work for users. As Vercel tech manager Brian Emerick noted, “Soon there will be more agents running around a company than people.”

The collective rush to build Claw‑style agents reflects a consensus that the next growth frontier for AI lies in system‑level automation and monetizing actual work output.

Reference Links

https://x.com/itsPaulAi/status/2027128959717314772?s=20

https://x.com/ManusAI/status/2023412626428932494?s=20

https://x.com/NotionHQ/status/2026356464537972900

https://x.com/cryptopunk7213/status/2026750915655373197?s=20

https://x.com/perplexity_ai/status/2026695550771540489?s=20

Code example

[1]https://x.com/itsPaulAi/status/2027128959717314772?s=20
[2]https://x.com/ManusAI/status/2023412626428932494?s=20
[3]https://x.com/NotionHQ/status/2026356464537972900
[4]https://x.com/cryptopunk7213/status/2026750915655373197?s=20
[5]https://x.com/perplexity_ai/status/2026695550771540489?s=20