Team Debt: The biggest trap facing startups today, and how to reduce the risk of it accumulating.

The concept of technical debt is familiar to most founders. In an AI world, it's also becoming less relevant - the interest rates on tech debt have plummeted in recent months, and paying it down can be done quicker than ever. But startups are about to face a new and more expensive kind of debt, one that can't be refactored away.

Team debt is what you accumulate when you build an organisation around assumptions that are no longer true, or will stop being true in the near future, and it is one of the biggest potential traps facing any startup or scaleup today.

Why a new form of debt is here

For decades, if you needed more output, you hired. As the business became more complex, you added process. If context was being lost within that process, you added meetings. And if a team was overwhelmed, you opened another role, hire again, and the cycle would continue.

This made sense when humans were the only operators inside a business. It makes much less sense now.

AI fundamentally changes the shape of what good work looks like. Agents are beginning to own workflows rather than just sit around them, and the gap between what agents can do and what companies are asking them to do is widening every month. If you accept that the future of an organisation is humans and agents working in collaboration - and you should, because it is - then the structure of teams must change to get the most out of that collaboration.

But the majority of companies aren't changing. They're hiring normal teams, building normal processes, and adding AI tools on top. Each one of those hires goes into a job where there's a pre-baked assumption of how work should be done. But that assumption, in many cases, is now wrong.

Every layer added into this structure is a new layer of team debt - and each layer will need paying down eventually.

How teams actually work now

In engineering - which is furthest along - individual humans are already directing upwards of twenty agents in parallel. But they aren't speaking to twenty different agents. Instead they speak to one orchestrator agent that directs the rest (this is a slight oversimplification, but its directionally correct and will do for now).

These orchestrator agents are beginning to look a lot like employees, and before long, it won't be abnormal to have orchestration level agents acting just like a normal team member, taking complete ownership of sizeable chunks of work and calling on agents or other tools to help complete that work.

But if agents are part of the team, you can't run the team on the same processes you built for humans. Asking an agent to join a morning standup and speak for three minutes makes no sense when it can generate a status report every hour without any drop in productivity. If an agent needs to direct another agent, having them send an email or a Slack message is absurd when they can write code and execute a script instead. Current processes assume that the office is closed between 6pm and 8am, but agents work 24/7.

In a real AI-native business, agents already outnumber humans quite drastically - a five person startup might have fifty or more agentic employees running at any given time. This is not a future prediction, it's true now. At that ratio, you don't bolt agents onto human processes. You design process for agents first, and add human process where it's needed on top to allow humans to keep up.

It is outright impossible to bring agents into an existing team structure and expect to leverage them fully without building new process from the ground up. And new process from the ground up means that almost every human role changes shape. Many must change so drastically that the old role shouldn't exist at all - and this is where team debt sits.

Unnecessary complexity

Forward deployed engineers are one of the clearest examples of where complexity causes issues. An FDE exists because a company's product and process are complex enough that a human needs to sit between the product and the customer to enable value delivery.

That might be a defensible decision today, but it is creating a compounding problem - every layer of process built to support that complexity is a layer that agents can't easily work within, and every hire made to manage it adds more team debt on top.

What feels like a strategic advantage today - "our product is so sophisticated it requires human-led deployment" - becomes an incredible bottleneck in the near future.

An AI-native competitor won't necessarily aim to replace the FDE with an AI agent doing the same job. They'll remove the conditions further upstream that made an FDE necessary at all, designing their product so that AI can handle the deployment and the onboarding without requiring a human in the loop.

Companies who win will optimise for the highest possible surface area where AI can be leveraged.

When I ask founders why agents can't be leveraged more, the answer is almost always "because of the complexity involved."

When you ask "why does that complexity need to exist", the answer rarely makes sense when said aloud. It's often a bit of a penny drop moment.

Team debt is different from overhiring

Overhiring is bad, but it is not team debt. Overhiring means you have too many people, but team debt means you have people in roles that shouldn't exist in their current form, either now or in the near future.

One almost certainly leads to the other, but team debt can show up in teams and companies who hire in very small numbers, and hiring too many people is not the only cause.

How to manage team debt

I'm not sure if you can entirely prevent team debt. To do so would mean being entirely accurate with all of your predictions about the upcoming few years. However, you can reduce the risk of it accumulating.

I have heard AI be described as "Ozempic for hiring".

In other words, AI, when utilised properly, should suppress your appetite to hire. Unless the company is genuinely starving for capacity or capability that doesn't exist inside the current team, you probably shouldn't be adding to it. Instead, you should be looking at the outcomes that need to be achieved, and ask whether there is a workflow to be created where an agent (or team of agents) can get you closer.

When you do hire, the bar needs to be extreme - both in skill and adaptability. Humans in your company need to be better than not just one AI agent, but 100, or 1000. They need to be doing work that is fundamentally important, that can't justifiably be removed from the company entirely (even with first principles thinking), and that agents can not do well. There is still a lot of this type of work that must exist, and humans are still important.

Those humans should also understand how to work with agents in a collaborative way. Not just instructing their own agents, but adapting the way they work to be more easily leveraged by agentic team members, and being willing to lean into this somewhat daunting and unproven team structure.

Although these teams are likely to be smaller (or, include less humans at the very least) - I want to be clear that the goal is not simply to be smaller. It's to build teams on structures where AI leverage is easy to come by, and avoid process where it isn't.

Nobody has all the answers. But the first step toward the right answer is to let go of those that you already know are wrong now.