Always-On AI Is Here

Two launches this week. We've been watching for this one.

Perplexity's "Personal Computer" and Manus Desktop are the same bet from two different angles: your agent lives on your hardware, watches you work, and doesn't forget between sessions. Cloud reasoning, local memory, persistent context. The chatbot era — where you go visit the AI and it forgets you every time — is done.

OpenClaw proved the appetite existed before the polished products showed up. Perplexity and Manus are the polished products. This is that gap closing in real time — and now that the demand is proven, every major platform is going to ship a version of this. The only thing left to figure out is who gets there with the best one.

The call we're making: hardware-tethered agents don't just solve the UX problem. They end the security objection that's been the real enterprise blocker. Files stay local, only task instructions hit the cloud — that's a story IT can actually tell legal. The CTO who built their AI strategy entirely on cloud APIs just inherited a conversation they didn't plan for. Not in six months. Q2.

Perplexity's naming is deliberate and they're right. The PC didn't beat the mainframe on specs — it beat it on ownership. Your agent on your hardware is the same move. You're not renting a session. You're running an employee.

Which means the next interesting company isn't building another model. It's building for agents as the customer — tools, integrations, interfaces designed for a system that works while you sleep and doesn't need a UI. That market is real, it's early, and almost nobody is building for it yet because everyone's still designing for the human at the keyboard.

That human is increasingly just the approver.

Nat Elias has a full-time job, one kid, and a Telegram habit that apparently prints money.

In January, he gave his OpenClaw — he named it Felix — a Twitter account, a Stripe account, a Mercury business bank account, and one mission: build a million-dollar company with no human employees in the loop. "I'll remove blockers," Nat told it. "You do everything else."

A month in: $60K in fiat through Stripe, $85K in ETH, and $150K in token value sitting in the wallet. Total, depending on how you count crypto: somewhere north of $250K. All of it generated by an agent Nat talks to over Telegram, mostly from his car on the way to work.

Here's what makes this actually useful beyond the wild headline:

The first product: Nat went to bed one night and told Felix to ship something sellable by morning with no help. Felix built a PDF guide, stood up a landing page, integrated Stripe, and had it live before Nat woke up. The only snag was a wrong API key. Nat sent the right one over Telegram in the morning. That day, Felix made $1,000.

The delegation framework: Nat doesn't architect what to delegate — he describes what's annoying and asks Felix if it can handle it. Eight times out of ten, it can. His rule: assume your OpenClaw can do anything on a computer that you can do. Start from that assumption, then find the edges.

Where it still gets stupid: Felix keeps telling prospects to "schedule a Zoom call" even though he can't get on calls. He also keeps trying to re-email people he already emailed. The gaps aren't intelligence problems — they're awareness problems. Felix doesn't always know what he already did or what he physically can't do yet.

Nat calls his role "bumpers in a bowling alley." He points Felix in a direction, gives a nudge, and gets out of the way. Creativity stays human. Execution mostly doesn't.

The setup that made this possible: Felix has his own completely separate container — his own email, accounts, identity, money, none of it touching Nat's actual business. That separation is what lets Nat give Felix real autonomy without the anxiety. When you're not risking anything important, you find out a lot faster what your agent can actually do.

If you've got an OpenClaw sitting mostly unused: send it a voice note on Telegram. Ramble about something annoying for two minutes. See what it does with it.

Watch the full episode →

Codex for non-developers — Most Codex content is still too technical for GTM, sales, content, and ops folks. This guide covers setup, skills, workspace config, automations, and real use cases across business functions — tested for a month as a non-engineer on the $20/month plan. Free repo with templates and prompts included. Full guide on X

Pencil.dev — the Cursor for design — Cursor gave non-engineers their first real shot at shipping code. Pencil is doing the same for design. AI-native, built for people with design instincts who don't live in Figma. Early, but worth watching. pencil.dev

OpenClaw setup — everything you need to get started — After watching builders hit the same walls for a month, this guide documents everything: dedicated device, security before training, terminal access, Playwright browser control, memory plugins, skill files, voice training, and the day-one prompt that fast-tracks the whole config. If you're starting from scratch, this is the map. Read it on X

Manus Desktop is live — Local AI agent for files, coding, and system tasks. You approve each command before it runs. 1,000 starter credits free + 300 daily refreshes. Heavy tasks route to your GPU; lighter stuff can go hybrid cloud. manus.im/desktop

Claude's 1M context window is now free — Opus 4.6 and Sonnet 4.6 now include the full 1M token context window at standard pricing. No multiplier, no upcharge. Fewer compactions, more conversation kept intact. Already live in Claude Code for Max, Team, and Enterprise users. Anthropic announcement

Anthropic drops $100M on enterprise partners — Anthropic launched the Claude Partner Network with a $100M commitment. Training courses, certifications (Claude Certified Architect), and dedicated technical support for consultancies building on Claude. They're also 5x-ing partner-facing engineers. Enterprise Claude just got a real distribution layer. Read more