I thought I understood AI agents, but something has changed over the past few weeks.
They are starting to feel less like software you use and more like people you hire.
I saw it most clearly at Anything, a 15-person company that says they are now running more than 100 agents to run their company with more than a million users. I’ve been getting close to the team and it’s going to change how companies work.
These are not chatbots waiting for prompts. Their agents have names, jobs, memory, tools, and their own computers in the cloud. It’s weird.
Most people are still using chatbots that wait for you. You ask a question and then you get an answer. You control the work end to end.
Tools like Skydive and Grokbot invert that. The agents keep going even when you close your laptop.
You’re not really prompting anymore. It looks more like hiring. It’s weird.
Stop prompting
The lines are blurring between what we call an agent and what is a coworker. Most agents are simply online. They live in your browser.
New agents live where you live. They have computers, they use the browser, they can code and share files. They sign into the same tools you already use.
They stop needing to be prompted.
If you think about these products like ChatGPT with extra buttons, you’ll be disappointed. If you treat it like someone you just added to the team, the product starts to make sense.
Skydive allows you to add your agent to slack. They get real profile pictures and can respond to you inline. This is very cool.
They don’t just respond tho, they use all their tools to get real work done.
They all have names too:
- Emma works in customer support.
- Grace investigates bugs.
- Canary watches production.
- Polly makes ad creative.
- Rice does outbound sales.
- Otis turns conversations in Slack into Linear tickets.
- Justin orders snacks for the office.
The system flows. Each agent is really specific and good at their individual job and the company runs with or without humans.
When I hire someone, I do not expect to walk over to their desk every fifteen minutes and give them another prompt. I give them a job. They think about it and they try to make progress on it every day.
They learn where things are. They get accounts. They watch what is happening. They remember what happened yesterday. They ask questions when they get stuck. Eventually, if they are good, I stop thinking about most of the work they are doing.
That’s how Skydive works.
How Anything ended up with 150 agents
Anything wasn’t really in the business of building AI employees. Anything is an AI app builder, like Lovable.
The founders were running Anything, but they were running into the same problems a lot of startups that I talk to run into.
There’s never enough people to do the things that need to get done.
More customers needed help. Fixing bugs becomes a full time job while the AI industry is moving faster than ever. More things needed monitoring and they were trying to play catch up with Lovable and Replit.
Companies used to hire people to try to deal with this. They decided to build a product that made it simple to create new AI Agents.
Then something interesting happened. People inside the company kept finding more jobs.
Justin ordering snacks for the company started as a joke, but then they realized that it just worked and now everyone can order snacks by DMing an agent in slack. They didn’t start out to build an AI snack ordering product. It just happened.
This is how you end up with 150 agents. Some of them are more fun, some are focused on operations. Many are focused on operations and of course some are focused on growth.
This changes how you build an agent
This is the part I got wrong. When someone gives me access to a product, I usually try to make it as powerful as possible. I give it access to all of my tools. This is the logical approach, but over time most of those products start to break down.
Skydive has focused on “narrow agents”. These agents are way more specialized. They roughly map to roles at the company. Emma isn’t going to order you lunch, but she will reply to customers better than any other agent at the company.
Each agent has a job, which makes the whole thing easier. It has the information you need. It knows what software it needs. It is more secure and it has better judgement.
Multiple models for agents may work out over time, but this one seems to be getting the most traction. Especially as AI costs rise on frontier models.
Magic
I got pretty deep into their product. I realized that they do iMessage as well (I’m sure you’ve seen the Instinct hype)
I tried it out and it was magical. The idea of using the same agent everywhere I go feels right.They have a CLI so you can use it next to Claude Code, Codex, or Cursor.
Maybe texting with agents is the last interface? Maybe ChatGPT was just a fad.
I tried putting multiple in a group chat and the results were mixed. The agents were good at working together, but they don’t always arrive at a final result. Sometimes they’re too deferential to each other. This still needs work but it feels like a glimpse into the future.
What is left?
Agents still need your judgement. You can’t leave them alone and expect to have a huge company created when you come back in 6 months. But it does make it easy to remove a ton of work from your plate.
These agents have already started to remember things that even I didn’t expect. They can recall conversations that we’ve had and will remind me of things i’ve asked for.
This is the part that feels different.
For the last few years, the big question in AI was which model. There’s an entire industry of evals trying to assess what a model can do when you send it a prompt.
The next question si what it can do when you stop prompting it.
Maybe “AI coworker” is still t he wrong word. Maybe all of this is a bubble and it breaks in ways we have not seen yet.
But after spending the last few weeks working within systems like this, I am increasingly convinced that the weird part is not that ANtyhing has more than 100 agents.
The weird part is that, pretty soon, that might not seem weird at all.
Paid partnership with @ANYTHING