An Agent That Runs a Whole Company: What Should You Delegate?
Andon Labs opened Pion to the public this week, a platform where AI agents run businesses end to end. Here is what founders should hand over now, what to keep, and how to pilot one without risking the company.
What Andon Labs actually shipped this week
On September 14, Andon Labs opened up Pion, a cloud platform where AI agents run continuously and handle the operating work of a business. You don't script tasks. You give a high-level directive to a supervisor agent called Andonos, and it coordinates the rest.
The announcement hit the Hacker News front page within hours and pulled more than 300 comments. Most were not about the product. They were about the claim.
"Designed to run any company autonomously" is a big sentence. Founders reading it fall into two camps. One sees a labor cost line going to zero. The other sees a demo that will quietly break the first time a supplier misses a delivery.
Both are partly right. The useful question isn't whether agents can run a company. It's narrower: which parts of your company are already shaped like something an agent can hold, and which parts would fall apart the moment nobody was watching?
That question has a real answer, and it's the same answer whether you're evaluating Pion or building your own agent stack.
The vending machine that lost money for a year
Andon Labs has a track record here, and it's more interesting than the headline.
In early 2025 the team put a physical vending machine in Anthropic's office and let an agent run it. The agent lost money. Not a little. It made poor pricing calls, got confused by the messiness of a physical environment, and struggled with the kind of judgment a bored teenager develops in a week behind a counter.
Then the underlying models improved. By late 2025 the same setup turned profitable. In April 2026 the team scaled up to harder things: Andon Market, a retail store in San Francisco, and Andon Cafe in Stockholm.
So the fair read on the technology is not "agents can run companies." It's "agents got from unprofitable to profitable on a vending machine in about nine months, and the ceiling is moving."
That tells you the capability is real but recent, the failure modes are about physical and social unpredictability rather than reasoning, and whatever you decide today about delegation will be wrong in six months. Plan for revision, not a permanent answer.
Why "runs the company" is doing a lot of work in that sentence
Every company is really two companies stacked on each other.
The first is a machine. Orders come in, invoices go out, tickets get answered, inventory gets reordered, the standup notes get written. This part has rules. Sometimes undocumented and often stupid, but they exist, and someone could write them down.
The second is a set of bets. Which customer segment to chase. Whether to raise now or in March. Whether the thing your biggest account keeps asking for is a roadmap item or a distraction. Whether to fire the salesperson everyone likes.
Agents are getting good at the machine. They are not close on the bets, and the reason isn't reasoning ability. The bets require accountability. Somebody has to own being wrong in a way that changes their behavior next time.
When a vendor says "runs your company," they mean the machine. That's still a lot. For a solo founder the machine might be sixty percent of the calendar. Getting most of that back is worth serious money. Just don't confuse it with the other thing.
What agents handle well right now
Based on what's shipping in 2026, the reliable zone looks like this.
Work with a clear definition of done. Reconciling Stripe payouts against your ledger. Turning a support transcript into a bug ticket with steps to reproduce. Drafting the weekly investor update from your metrics dashboard.
Work that is high volume and low variance. Screening inbound leads against your ICP criteria. Categorizing expenses. Chasing overdue invoices on a schedule. Monitoring competitor pricing pages and flagging changes.
Work you already hate and do badly. This one matters more than people admit. The tasks you procrastinate on are usually the ones where an agent's mediocre output beats your actual output, because your actual output is "next Tuesday, maybe."
Research with a verifiable end state. Pulling four quarters of a public competitor's filings. Building a list of every integration partner in a category.
Notice what these share. Each has an observable result you can check in under two minutes. That check is the whole game. If verifying the work costs as much as doing it, delegation saves you nothing.
Your AI co-founder is ready when you are.
Foundra turns everything in this article into an actual plan. Validation, customers, pricing, launch. In one place, in your voice, in an afternoon.
Get started→$39/month. Cancel anytime.
What stays on your desk
Keep anything where being wrong is expensive and the wrongness shows up late.
Pricing changes. Hiring and firing. Anything that touches a contract. Public statements in your company's voice. The decision to kill a product line. Conversations with your five largest customers, even the boring ones, because those relationships are an asset and agents don't build relationships, they process interactions.
Also keep early customer discovery. Not because an agent can't run interviews, but because the value of those calls is what changes in your head while you're on them. Outsourcing that is like paying someone to go to the gym for you.
Here's the part founders get wrong. They delegate the machine work and then never look at it again, and six months later they can't answer a diligence question about their own churn because a process they never watched has been quietly producing numbers nobody reconciled.
If you map out which functions you're keeping and which you're handing off, put it somewhere structured rather than in your head. A spreadsheet works. So does Notion, or a planning tool like Foundra that walks first-time founders through mapping operations alongside the rest of the business plan. The format matters less than the fact that it's written down and revisited.
A three-question filter before you hand anything over
Run every candidate task through these.
1. Can I describe done in one sentence? "Reconcile the September Stripe payouts against QuickBooks and flag anything over a $5 dollar variance" passes. "Improve our onboarding" does not. Vague directives produce confident nonsense.
2. What does failure cost, and when do I find out? A misfiled expense costs twenty dollars and surfaces at tax time. A mispriced enterprise quote costs the deal and surfaces never, because the prospect just stops replying. Delegate the first kind first.
3. Do I have a check that takes less than two minutes? A dashboard number, a spot check of three records, a diff against last week. If you can't build the check, you're not delegating. You're hoping.
Tasks that pass all three go to an agent this month. Tasks that fail question three go on a list to revisit once you've built the check. Tasks that fail question two stay with you regardless of how tempting the automation looks.
Most founders can find eight to fifteen tasks that pass all three on the first pass. That's a real week of time back.
How to pilot one without betting the company
Start with a shadow run. Pick one process, let the agent do it in parallel with whoever does it now, and compare outputs for two weeks. You're not measuring whether the agent is impressive. You're measuring how often it's wrong and how the wrongness looks. Random errors are manageable. Systematic drift in one direction is what bankrupts you quietly.
Give it the narrowest credentials that work. Read-only wherever possible. Separate API keys per agent so you can revoke one without taking down everything. A spending cap on anything that moves money, low enough that the worst case is annoying rather than fatal.
Log everything and read the logs for the first month. Not skimming. Reading.
Then set a review date sixty days out with three numbers: hours saved, error rate, and cost. If you can't fill in all three on that date, the pilot failed regardless of how good it felt.
One more thing. Write down what you'd need to see to shut it off, before you start. Sunk cost hits automation projects hard, because turning one off feels like admitting you fell for a demo.
Key takeaways
- Andon Labs opened Pion on September 14, 2026. Its earlier agent went from losing money on a vending machine in early 2025 to profitable by late 2025.
- Every company is a machine plus a set of bets. Agents are getting good at the machine. Keep the bets.
- Delegate work with a one-sentence definition of done, low and fast-surfacing failure cost, and a check that takes under two minutes.
- Hold onto pricing, hiring, contracts, public voice, top-customer relationships, and early customer discovery.
- Pilot with a shadow run, read-only credentials, spending caps, and a sixty-day review.
- Write your shutoff criteria before you start.
FAQ
Is Pion ready to run a real business without supervision?
Not for most companies. Andon Labs has run its own retail and cafe operations with agents since April 2026, which is real evidence, but those are tightly scoped businesses the team designed around the technology. Treat it as a strong operations layer that needs a human owner, not a replacement for one.
What's the first thing a solo founder should automate?
Whatever you've procrastinated on twice this month that has a checkable output. Usually invoice chasing, expense categorization, or turning support conversations into tickets.
How do I know if an agent is quietly making things worse?
Systematic drift is the signal. Spot check a random sample weekly for the first month and look for errors that all lean the same direction. Random mistakes are normal. Consistent bias means the agent has learned the wrong rule and will keep applying it.
Does using agents this early make my company look unserious to investors?
The opposite, if you can explain your controls. Investors in 2026 expect small teams to do more. What worries them is a founder who can't describe what the agents touch, what the failure modes are, and who verifies the output.
What about agents handling customer conversations?
Fine for triage and first response on common issues. Not for your largest accounts, renewals, or anyone already upset. Getting those wrong shows up in churn, and churn surfaces too late to correct.
You just read the theory. Ready to build the thing?
Foundra is your AI co-founder. It turns an idea into a validated business plan, a go-to-market, and your first 10 customers. In an afternoon, not a semester.
$39/month. Cancel anytime. Works in 20 languages.