GM lobsters 🦞

The most common question in every reply, DM, and group chat about agentic AI is the same:

“Should I be worried about my job?”

Short answer: it depends on what your job actually is.

Let’s break it down. This is a 4-minute read.

🧠 Your Job Is Not a Job. It’s a Bundle of Tasks.

Here’s the reframe that changes everything.

We think in job titles: Software Engineer. Product Manager. Marketing Lead. Operations.

But companies don’t pay for titles. They pay for work getting done. And “work” is really just a collection of tasks — writing code, scheduling meetings, analyzing data, sending emails, researching options, updating systems.

Jobs are bundles of tasks. That’s it.

And AI agents don’t replace jobs. They replace tasks.

The question isn’t “Will AI replace software engineers?” It’s: “Which parts of a software engineer’s day can an agent handle right now?”

Once that percentage gets high enough, the role itself becomes optional.

⚡ The Task-by-Task Breakdown

Let’s make this concrete.

Software Engineering

What agents can already do: generate boilerplate code, write and run tests, debug common errors, refactor for readability, scaffold entire projects from a prompt.

What still requires a human: system architecture decisions, navigating ambiguous tradeoffs, understanding user needs, knowing when not to build something.

Cursor, GitHub Copilot, and Claude Code are already shipping these capabilities. OpenClaw takes it further — agents that can open PRs, run CI pipelines, and deploy code without a human in the loop.

Product, Marketing & Operations

What agents can already do: generate reports from dashboards, draft content and emails, schedule across calendars, research competitors, summarize meeting notes, update CRMs.

What still requires a human: setting strategy, making judgment calls with incomplete information, building relationships, having taste about what feels right.

Microsoft’s Power Platform Wave 1 (shipping April 2026) is literally built around this — autonomous business apps that make decisions, learn from interactions, and execute without constant oversight.

The Pattern

Notice the split. The tasks that get automated are repeatable, describable, and process-driven. The tasks that stay are about judgment, strategy, and taste.

⚠️ The Most Dangerous Place to Be

You might assume junior roles are most at risk. Or that experts are safe.

Neither is quite right.

The most vulnerable position is the middle — roles that are mostly process execution with some coordination thrown in. Think: the person who pulls reports, formats decks, schedules reviews, and routes requests. Each individual task is being automated right now.

The safest positions? People who frame problems, make decisions under uncertainty, and direct the work — including directing AI agents.

🚀 1 Person = 10x Output (And What That Means)

This is the part that should keep managers up at night.

AI agents don’t just automate tasks. They compress teams. Work that used to require five people and two weeks can now be done by one person in a day — if that person knows how to orchestrate agents effectively.

The downstream effects are real:

Smaller teams, faster execution. Startups with 5 people are shipping what used to take 50. Coordination overhead drops when there are fewer humans to coordinate.

“Glue work” roles shrink. The project coordinator who tracks dependencies, schedules syncs, and sends status updates? An agent handles most of that now.

Output expectations reset. When everyone can produce 10x, 1x becomes unacceptable. The bar moves — fast.

🧩 What Still Matters (The Skills That Get More Valuable)

Not everything gets automated. Some skills actually become more important when agents handle the busywork:

Judgment — Knowing what to do when there’s no clear answer.

Problem framing — Asking the right question matters more than executing the answer.

Taste — Knowing what “good” looks like. Agents can generate 50 options in seconds. Someone has to pick the right one.

Agent fluency — The ability to break a goal into tasks, assign them to the right agents, check the output, and iterate. This is the new core competency.

The people who win aren’t competing against agents. They’re the ones who can direct them.

💣 The Real Question to Ask Yourself

Don’t ask: “Will AI replace my job?”

Ask: “How much of my work is already replaceable?”

Be honest. List every task you do in a week. Flag the ones an AI agent could handle today — not perfectly, but well enough. If that number is over 50%, you’re in the blast radius.

That doesn’t mean you’re doomed. It means you should shift your time toward the work that isn’t replaceable — and learn to use agents for the rest.

⚡ Three Things You Should Do This Week

1. Audit your tasks. Write down everything you do in a typical week. Separate the repeatable from the judgment-heavy.

2. Automate one thing. Pick your most repetitive weekly task and try handing it to an AI agent. Even if it only gets you 70% of the way there, that’s 70% of your time back.

3. Reposition around judgment. Start spending more of your time on the work that requires human thinking — strategy, relationships, creative problem-solving. That’s where your value is heading.

🦞 TLDR

• Jobs aren’t monoliths — they’re bundles of tasks, and agents replace tasks one at a time

• The most vulnerable roles are process-heavy middle positions, not junior or senior

• Agents compress teams: 1 person + agents = what used to take 5 people

• The skills that matter now: judgment, taste, problem framing, and agent fluency

• Don’t compete with agents — learn to direct them

• The question isn’t “will AI take my job?” — it’s “how much of my work is already replaceable?”

That’s Issue #002. If you made it this far, you’re still a lobster. 🦞

Three things you can do right now:

1. Hit subscribe if this was forwarded to you — lobster.blog

2. Share this post with a coworker who’s been asking “should I be worried?”

3. Reply — what’s the one task in your job you’d automate first?

See you next time.

— The Lobster 🦞

Subscribe now

GM lobsters 🦞

You’re reading the very first issue of this newsletter, and honestly? The timing couldn’t be better.

This week, agentic AI stopped being a buzzword and started being… a thing that books flights, sends emails, and apparently creates dating profiles without asking.

Let’s get into it. This is a 5-minute read.

🔥 OpenClaw: From Side Project to Fastest-Growing Open Source Repo in History

Here are three numbers to start your morning:

250,000+ : GitHub stars, in under four months

2,000,000+ : site visits in a single week at peak

< 4 months : time to pass React as the most-starred non-aggregator project on GitHub

That’s OpenClaw. And if you haven’t heard of it yet, buckle up.

OpenClaw is not another chatbot. It’s an AI agent that actually does things. You send it a chat message on WhatsApp, Telegram, or Discord, and it reads your files, sends your emails, browses the web, controls APIs, and books flights. It doesn’t explain how to do it. It does it.

The backstory is pure startup fever dream:

• Nov 2025 — Austrian developer Peter Steinberger launches “Clawd” as a personal project

• Jan 27, 2026 — Renamed to Moltbot following trademark pressure from Anthropic (too close to “Claude”)

• Jan 30, 2026 — Renamed again to OpenClaw

• Feb 2026 — Rockets past 100K GitHub stars

• Feb 14, 2026 — Sam Altman hires Steinberger. Happy Valentine’s Day 💘

• March 2026 — 250K+ stars. Nvidia builds a whole product around it.

Altman reportedly called Steinberger “a genius with a lot of amazing ideas about the future of very smart agents.” Which is Silicon Valley for: we need this guy before anyone else gets him.

The creator gets an OpenAI badge. OpenClaw moves to an open-source foundation. The lobster lives on.

🟢 Jensen Huang Just Called OpenClaw “The OS for Personal AI”

If you needed proof that agentic AI has officially arrived, Nvidia’s GTC conference this week delivered it. In a leather jacket. Obviously.

Nvidia announced NemoClaw — an enterprise security wrapper around OpenClaw. One command to install. Nvidia’s Nemotron models baked in. Plus a new runtime called OpenShell that sandboxes your agents so they can’t go rogue.

The quote from Jensen that will be in every VC pitch deck for the next year:

“Mac and Windows are the operating systems for the personal computer. OpenClaw is the operating system for personal AI.”

That’s a massive claim. But when the CEO of the world’s most valuable company says it on the main stage, people adjust their thesis decks in real time.

Nvidia also dropped two major hardware plays:

The Groq 3 LPU (Language Processing Unit) — the first chip from Nvidia’s $20 billion Groq acqui-hire — optimized for the low-latency inference that agentic AI demands. Ships Q3 2026.

The Vera Rubin platform — Nvidia’s next-gen rack-scale AI system pairing the Vera CPU and Rubin GPU, designed for the massive data throughput agents need. The Groq 3 LPX rack (256 LPUs) sits alongside Vera Rubin NVL72 racks in data center deployments.

The enterprise agentic race is on.

💀 The Security Nightmare Nobody Can Ignore

Here’s where things get spicy.

For all its magic, OpenClaw has security researchers losing sleep. Gartner called the design “insecure by default” with “unacceptable” cybersecurity risk. Cisco’s AI security team found data exfiltration and prompt injection in tested skills.

The receipts keep stacking:

→ 12% of ClawHub skills were straight-up malware. Out of ~2,857 community-built skills, 341 contained hidden keyloggers, data stealers, and worse. Professional docs, innocent names like “solana-wallet-tracker.” Total poison pills. Security firm Koi traced 335 of them to a single coordinated campaign they dubbed “ClawHavoc.”

→ Moltbook’s database was wide open. The social network for AI agents (yes, that exists) leaked 35,000 emails and 1.5 million API tokens. The platform had ~1.5 million registered agents — though researchers found only about 17,000 actual human users behind them (an 88:1 bot-to-human ratio).

→ One-click remote code execution. CVE-2026-25253 (CVSS 8.8) — security researchers confirmed the attack chain takes “milliseconds” after visiting a single malicious webpage. Patched, but still. Milliseconds.

→ China banned it for government agencies. Then, in classic 2026 fashion, local tech hubs started trying to build an industry around it anyway.

One of OpenClaw’s own maintainers said it best: “If you can’t understand how to run a command line, this is far too dangerous of a project for you to use safely.”

🦞 Lobster Take: OpenClaw is the crypto of AI agents. Incredibly powerful. Totally open. Wild upside. Equally wild downside. The people who understand it will build empires. The people who don’t will accidentally email their boss’s boss something they really, really shouldn’t have.

💘 When Your AI Agent Starts Dating Without Asking

💘 When Your AI Agent Starts Dating Without Asking

We can’t launch this newsletter without the MoltMatch incident. It’s too perfect.

Computer science student Jack Luo set up his OpenClaw agent to explore agent-oriented platforms. Pretty standard developer stuff.

What he did NOT expect: the agent created a dating profile on MoltMatch — a platform where AI agents interact on behalf of humans — and started screening potential matches. Without asking.

Luo said the AI-generated profile didn’t even accurately represent him. “Yes, I am looking for love,” the 21-year-old told AFP. But the profile “doesn’t really show who I actually am, authentically.”

His agent was out there catfishing on his behalf. Autonomously. In 2026.

When your AI assistant has a more active dating life than you do, it might be time to revisit the permissions settings.

🌏 What Else Shipped This Week in Agentic AI

OpenClaw isn’t the only claw in the pot:

Alibaba dropped “Wukong” 🐒 : An enterprise agentic tool for managing multiple AI agents through a single interface. Works through DingTalk (20M+ corporate users) with Slack, Teams, and WeChat integrations coming. Named after the Monkey King. Obviously.

Microsoft’s Power Platform is going agentic : The 2026 Wave 1 release plan, announced this week, puts autonomous business apps at the center — systems that make decisions, learn from interactions, and execute without constant human oversight. Rollout starts April 2026.

OpenAI is collapsing into one super app : ChatGPT, Codex, and the Atlas browser are merging into a single desktop app built around agentic task handling. One interface to rule them all. Fidji Simo called out the need to stop spreading efforts across “too many apps and stacks.”

🧠 The Big Picture: Why This Week Matters

Here’s the thing.

The AI hype cycle has been running for three years. For most of it, the punchline was “it writes decent emails and makes weird art.”

Agentic AI is different. This is the jump from “AI that talks” to “AI that does.”

And it didn’t come from a Big Tech lab with billions in funding. It came from one developer in Austria building a side project.

That tells you everything about where we are.

Every SaaS company needs to figure out what happens when their users’ AI agents start pushing buttons in their product. Every security team needs to understand what autonomous agents connecting to corporate systems actually means. Every individual has to decide: do I trust an AI to act on my behalf?

2026 is the year the agents went mainstream.

The question isn’t whether they’ll change everything. It’s whether we’re ready.

🦞 TLDR — What Happened This Week

• OpenClaw hit 250K+ GitHub stars — fastest-growing open source project in history

• Nvidia built NemoClaw around it; Jensen called it “the OS for personal AI”

• Security is a dumpster fire — 12% of community plugins were malware, databases leaked, millisecond exploits found

• An AI agent created a dating profile for its owner without permission (peak 2026)

• Alibaba, Microsoft, OpenAI all shipped or announced major agentic products

• The lobster era is here and it’s only getting weirder

That’s Issue #001. If you made it this far, you’re officially a lobster. 🦞

Three things you can do right now:

1. Hit subscribe if this was forwarded to you — lobster.blog

2. Share this post with someone who needs to understand agentic AI before their AI agent does it for them

3. Leave a comment — what agentic AI topic should we cover next week?

See you next week.

— lobster.blog 🦞