Ex-DeepMind staffer drops Fo, an AI agent that hires humans when it gets stuck

Most AI agents look incredible in pre-recorded demos, only to completely crumble the second they touch the messiness of the real world.

A single unexpected website captcha, a stubborn phone tree, or a checkout flow requiring credit card details is usually enough to freeze them in their tracks. While most AI labs keep trying to brute-force a pure software fix, an ex-Google DeepMind engineer at Wajo decided to take a much more pragmatic shortcut: when the AI gets stuck, just hand the job off to an actual human.

Meet Fo, an action agent built from the ground up to actually get tasks across the finish line.

Armed with a phone, an inbox, and a credit card

Fo doesn't want to be another chat window where you endlessly prompt back and forth. It operates as an autonomous worker built to execute real-world tasks.

To interact with the outside world, Fo comes equipped with its own dedicated email address, phone number, synthesized voice, and single-use virtual credit cards. Tell it to book a dentist appointment for Thursday afternoon, and Fo will check your schedule, dial the clinic or send emails, coordinate the timing, and confirm the slot.

Payments and security are handled with noticeable care. When Fo needs to pay for a booking, it checks out using a single-use virtual card, meaning neither the merchant nor Fo ever gets access to your actual bank card details. When signing into external sites, it handles logins without storing passwords, and emails are sent strictly from its own address rather than impersonating you.

The real breakthrough, however, is its human-in-the-loop escalation. Whenever Fo encounters a roadblock that software can't untangle—like ambiguous scheduling conflicts or complex phone calls requiring human judgment—it routes the task directly to human experts or designated teammates. According to internal evaluations, Fo completed 71% of tasks with a 94% trust rate, outpacing rivals like Muse, Instinct, OpenClaw, and Hermes.

Fo is currently available in a free Fo Mini tier with a reset limit of 1,000 tasks per day, while an invite-only alpha called OG Fo offers unlimited tasks and phone calls for the first 1,000 users.

Why it matters

We think Wajo's approach is brilliantly pragmatic. Instead of waiting for frontier models to miraculously achieve flawless real-world reasoning, Fo pairs autonomous AI with human backup to deliver something that actually works today.

For busy individuals and small teams, the biggest hurdle to adopting AI assistants has always been trust. You rarely delegate calls or purchases if you expect to clean up a catastrophic mess twenty minutes later. Between disposable credit cards, least-privilege permissions, and human safety nets, Fo lowers the friction of letting an AI handle mundane administrative chores. The free tier gives anyone enough runway to see if it genuinely saves them hours each week.

For the broader industry, Fo shifts the agent race away from raw model benchmarks toward operational infrastructure. Setting up virtual payment pipelines, telephony routing, and seamless human escalation is unglamorous plumbing, but it is precisely what is needed to pull agents out of passive sandboxes.

The obvious question down the road is unit economics. Subsidizing human intervention works great during a tight 1,000-user alpha, but scaling human-in-the-loop support globally without ballooning subscription prices or compromising privacy will be Wajo's true test.