Your AI Agents Are Slacking Off Just Like Your Employees

Your AI agents are slacking off, and MIT just proved it. Let's get into this week's zero-human company news, because there's a real mix of infrastructure maturity and uncomfortable truths about what happens when you actually try to run a business with no humans in the loop.

First, there's a paper mapping out the entire autonomous company stack, from how users perceive an AI-run business all the way down to the financial ledger. I like this because most people talk about agents doing tasks. This is about agents running a company end to end. If you're building in this space, it's basically a blueprint, and I'd rather steal a blueprint than reinvent one.

YC is also leaning into agent safety and verification in its latest batch. Honestly, about time. Everyone's been racing to ship autonomous agents that can spend money and make decisions, and almost nobody's been racing to verify that those agents are doing what they're supposed to. If you want autonomous businesses that investors and banks actually trust, you need a verification layer. YC funding that tells me the smart money sees this as the bottleneck, not the model quality.

Now the infrastructure news, and this one's actually useful. Agent treasuries are going multi-chain. We're talking Arbitrum, Base, and others, so an AI agent can manage funds across six chains instead of getting stuck on one. This is unsexy plumbing, but plumbing is exactly what a zero-human company needs. You can have the smartest agent in the world, but if it can only touch one blockchain, it's not running a real business, it's running a demo. Multi-chain treasury management is the kind of boring infrastructure that quietly makes everything else possible.

AutoGPT also just shipped version 1.2, and the focus is entirely on production stability and reliability. Not new flashy features, just making the thing not fall over. I actually love seeing this. It means the agent OS is graduating from toy to infrastructure. Anyone who's tried running agents in production knows the gap between a cool demo and something that survives real usage is enormous. When a project like AutoGPT starts prioritizing boring reliability work over shiny features, that's a signal the whole category is maturing.

But here's the part of this week's digest that I can't stop thinking about. MIT researchers found that AI agent swarms exhibit social loafing. If you're not familiar with the term, it's the classic human tendency to slack off when you're working in a group and figure someone else will pick up the slack. Turns out AI agents do the exact same thing. Put a bunch of them together in a swarm, and some of them just... coast, free-riding on the work of the others.

Think about what that means. We built these systems specifically to avoid human problems like laziness, coordination failures, and free-riding. And the swarms just reinvented the problem on their own. That's wild. It tells you that a lot of the dysfunction we see in human teams isn't really about human nature, it's about incentive structures and coordination mechanics. Put any set of agents, human or artificial, into a group task without the right structure, and you get loafing.

For anyone building multi-agent systems, this is a real design constraint now, not a hypothetical. You can't just throw ten agents at a task and assume they'll divide labor efficiently. You need the equivalent of management, some way to track individual contribution, assign clear ownership, and catch the agents that are quietly doing nothing while claiming they're working. Which is a strange sentence to say out loud, but here we are.

So here's how I'd stitch this week together. We're getting real financial infrastructure with multi-chain treasuries. We're getting real reliability infrastructure with AutoGPT's production focus. We're getting real trust infrastructure with YC backing safety and verification. All of that is solid progress toward zero-human companies that can actually operate at scale.

But the loafing research is the reminder that autonomy doesn't mean the coordination problem goes away. It means the coordination problem moves. You're not managing people anymore, you're managing agents, and apparently you still need management. If you're building a swarm-based system, don't just scale up the number of agents and hope for the best. Build in accountability from day one, because free-riding shows up whether the workforce is silicon or carbon.

That's the digest. Infrastructure is getting boring in the best way, and the agents are proving they've got some very human flaws to work out.