ClickUp Replaced Hundreds of Staff With AI Agents

Nine years. Hundreds of employees. Then ClickUp announced it is replacing most of them with AI agents.

That sentence would have read like science fiction in 2023. Today it's a TechCrunch headline and a board-approved strategy.

This week also brought the Pope's first encyclical on AI, a federal quantum bet with legal questions attached, and a bug-hunting arms race that security researchers say is already changing how exploits get built. The signal-to-noise ratio was high. We filtered it down to what matters for builders.

Below: 5 signals, 5 repos, 2 skills, 3 stack picks, and one Claude setup tip. Tuesday's issue. Let's go.

---

1. ClickUp Is Replacing Hundreds of Staff With AI Agents The nine-year-old project management startup is cutting its human workforce and deploying thousands of AI agents in their place. Not a pilot. Not a test. A structural shift. TechCrunch

2. The Pope's Encyclical Uses AI to Diagnose Power, Not Technology Pope Leo XIV released "Magnifica humanitas," his first encyclical. AI is the frame. The actual subject is concentrated power, eroding democracy, and a tech elite shaping the world for itself. TechCrunch · The Verge

3. The US Quantum Bet May Have a Legal Problem Washington placed a large bet on quantum computing infrastructure this week. Ars Technica reports the deal may have outrun the legal authority that funded it. Ars Technica

4. AI Is Rewriting the Security Exploit Playbook Attackers are using AI to find vulnerabilities faster. Defenders are doing the same. Wired maps the arms race and what it means for software security going forward. Wired

5. OpenAI Signs Brazilian News Partners OpenAI brought Grupo Folha and Grupo UOL into a content partnership that puts attributed Brazilian journalism inside ChatGPT. The media-AI licensing pattern keeps spreading. OpenAI

---

[Repo] gptme, Your agent in your terminal with local tools (writes code, uses the shell, browses the web). 4.3k stars and shipping fast. The closest open-source thing to a personal autonomous engineer.

[Repo] microsandbox, Secure, local, programmable sandboxes for AI agents. 6.3k stars. If you let Claude or Codex execute code on your machine, this is the cleaner way to gate it.

[Repo] Vision-Agents, Open voice and vision agents from Stream. 7.8k stars. Build agents that see and hear in real time using their edge network, free for hobby use.

[Repo] open-interpreter, Lets Claude and other models execute code, browse files, and run terminal commands locally. Bridges the gap between "Claude suggests a fix" and "Claude runs the fix."

[Repo] screenshot-to-code, Takes a screenshot of any UI and returns clean HTML/CSS/React. Pairs well with vision-based extraction workflows.

[Skill] Puppeteer for LLM pipelines, Most tutorials use Puppeteer for scraping. The real unlock is pairing Puppeteer screenshots with Claude vision for site auditing. Puppeteer captures, Claude reads, your pipeline decides. Runs in ~3 seconds per prospect at meaningful quality.

[Skill] Gmail OAuth with compose-only scope, Setting gmail.compose instead of gmail.send keeps a human in the loop on every cold pitch. The draft appears in your Gmail outbox for review before anything goes out. Cold-pitch quality stays high. One OAuth config change.

---

[MCP] Google Maps Places API v1, Pulls business name, address, website, and phone number for any search query and geography. Chain-filter it to skip franchises and aggregators. What's left is a clean prospect list of independent businesses with bad websites and real decision-makers.

[Tool] Vercel REST API, Deploy a single-file HTML site to a live preview URL in one API call. No CLI required. Useful anywhere you need to show a client what they'd buy before they sign anything. The free Hobby tier covers it at any prospecting volume.

[MCP] Anthropic Claude Haiku, At roughly $0.0001 per reply classification, Haiku is effectively free for labeling inbound email responses. Positive, negative, not-interested, referral request, Haiku reads each reply and routes it. Keeps your tracker clean without touching a paid tier for routine classification.

---

Tip: Anchor your vision extraction prompts with an explicit "return only data" rule.

Claude vision is strong at reading screenshots and inferring brand identity. Colors, voice, service categories, trust signals. But if your prompt explains the schema you want back, the model sometimes echoes schema comments as field values. You get "primary_color": "The hex code for the dominant brand color" instead of "primary_color": "#2B4C8C".

Fix: add one explicit line to your extraction prompt. Something like: Return populated field values only. Do not return schema descriptions, instructions, or placeholder text as values.

One line. Cleans the output completely. Apply this rule any time you feed Claude a structured schema in a vision context. The underlying pattern is documented in Anthropic's prompt engineering guide.

---

ClickUp's announcement is useful because it names the thing most companies are still talking around. The question stopped being "could AI do this job?" a while ago. This week it became "when do we stop paying people to do it?" The Pope's encyclical and ClickUp's layoff landed on the same day. That's not a coincidence worth ignoring.

---

This Friday: AI Website Agency for Small Businesses

The kit drops Friday. Here's the what. An AI pipeline reads a small business's existing website, extracts their brand identity, renders a 10x-better version of their site in their own colors and voice, and sends it cold with their name on it. The prospect sees the exact site they'd buy before any money changes hands. That's the mechanic that makes $497 setup + $99/mo close. Full stack, all 8 modules, every prompt, Friday.

---

ClickUp moved. The Pope wrote. Builders shipped. See you Wednesday with the next signal batch, and Friday with the full kit.

See you tomorrow.