How to Manage a Team When AI Agents Handle Most of the Work

The startup founders who managed the biggest teams five years ago aren’t managing the biggest teams anymore. The shape of the org chart is changing, and nobody’s really named the new problem yet.

Sam Kececi, founder and CEO of The Sentience Company, gave one of the cleanest framings I’ve heard on how this shift actually feels from inside a founder’s head. He used to be CTO of a company called Macro, where he managed 15 engineers. Today at Sentience — his AI startup that emerged from stealth in March 2026 — he manages 9 humans plus, as he puts it, “however many growing number of agents.”

The reduction from 15 humans to 9 isn’t the interesting part. The interesting part is what he said next.

The cognitive load didn’t reduce. It shifted.

“We’ve replaced the cognitive load of managing people now with the cognitive load of managing fifteen agents,” Sam told me. That’s the sentence that made me stop and rewind the tape. Because it names something almost every founder building with AI agents is feeling but hasn’t quite been able to describe.

Managing a human means: hiring them, onboarding them, delegating work, reviewing their output, giving them context, resolving conflicts between them and other humans, checking in on their growth, sometimes managing their emotional state. It’s expensive. Every one of those functions has real cognitive overhead.

An AI agent doesn’t need most of that. It doesn’t need onboarding in the same sense. It doesn’t need career development conversations. It doesn’t fight with other agents about credit — or if it does, that’s a bug you can fix.

But it needs something different. It needs prompting. It needs clear scope boundaries. It needs permissions — what data can it see, what actions can it take, who does it escalate to when it’s uncertain. It needs review — you can’t just deploy an agent and forget it, because the failure modes compound. It needs coordination — if you have five agents working across five parts of your business, they need to know about each other, or they’ll duplicate work or contradict each other.

That’s a different muscle. It uses different attention. And the total load doesn’t go down — it just reshapes.

What actually replaces middle management

Here’s the useful piece for anyone building an AI-native team today.

The middle-management function that AI agents actually replace cleanly is what Sam calls being a “glorified information router.” At Macro, he described his own experience as a “switchboard operator” — sitting between sales and engineering, engineering and product, taking a piece of information from one team and handing it to another. It was, in his words, a “toll booth in the loop.”

That work — routing information between people, translating context, answering the same question five times to five different stakeholders — is exactly what an AI agent trained on your knowledge and given the right permissions can do. And that’s a real subtraction from your day. Sam described being able to be fully present in a call because his Sentience was, in parallel, answering three questions from his team on Slack. The interruption cost of middle-management work is what gets removed.

What doesn’t get removed: strategic decisions, taste calls, judgment about which agent gets to touch which situation, the human moments that build trust with employees and customers. Those still land on the founder. And they land with more weight, not less, because the routing work that used to distract from them is gone.

The hire-and-fire ratio nobody’s talked about yet

If you’re hiring right now, the ratio you should be thinking about isn’t just “how many humans do I need.” It’s “what’s the humans-to-agents mix I want at 20 people? At 50?”

Some roles compress cleanly into agent workflows: information triage, meeting notes, first-draft communication, calendar coordination, initial-screen sales conversations, standard technical due diligence. Others don’t: setting strategy, building relationships with high-context customers, making the calls that determine what the company becomes.

The founders I’ve talked to who are running AI-native teams tend to be smaller than the equivalent five years ago — Sam’s team at Sentience is 8 people at the six-week mark, and they’re already running engineering standup through their own product. The team headcount matters less. The ratio matters more.

The other thing that changes: interview processes start to look different when you’re hiring for a shape that includes agents. Sam told me Sentience doesn’t do LeetCode interviews. They work with you. They hire for EQ plus IQ. Because when the technical baseline gets amplified by AI tools, the human differentiator moves toward judgment, taste, and how you collaborate under uncertainty. And you can’t test that with a coding puzzle.

There’s a real question buried in all of this that I don’t think anyone has answered yet. If the middle-management layer compresses to agents, what does the emotional labor of running a company look like? Because Sam still meditates ten minutes a day. He still has cofounder-adjacent conversations with his Chief of Staff. He still shows up to interviews in person when he can. The management burden shifted — but the human weight of running a company didn’t shift. If anything, it got heavier, because the strategic decisions are now more visible.

That’s the part I’d love to hear more founders talk about. Not the AI tools. The felt experience of running a team that isn’t just humans anymore.

---

Frequently Asked Questions

How is managing AI agents different from managing humans? Managing AI agents removes the interpersonal overhead — no hiring, onboarding, or career development — but adds a new load: scoping their work, defining permissions, reviewing output, and coordinating between agents. The total cognitive burden doesn’t decrease; it shifts to different muscles. Sam Kececi, founder of The Sentience Company, went from managing 15 humans at his prior startup to 9 humans plus a growing number of agents.

Do AI agents actually replace middle management? AI agents cleanly replace the “information routing” function of middle management — answering the same question multiple times, translating context between teams, and coordinating simple decisions. They do not replace strategic decisions, taste calls, or the human trust-building that senior roles carry. Sam calls the old middle-management work being a “toll booth in the loop.”

What roles at a startup compress well into AI agent workflows? Roles heavy on information triage, meeting notes, standard sales conversations, calendar coordination, first-draft communication, and initial-screen due diligence compress cleanly. Roles that require judgment under uncertainty, deep customer relationships, or strategic decision-making do not. The rule of thumb: if a role can be described as “routing information,” it’s a candidate for automation.

What’s the ideal humans-to-agents ratio for an AI-native startup? There isn’t a universal answer yet, but early AI-native teams like Sentience are running with roughly 1 human to 2-3 agents at the 8-person scale, and expect the ratio to widen as the company grows. Founders should think about this ratio deliberately: what work compresses to agents, and what work still needs the human?

How should hiring processes change for AI-native teams? Interview processes shift toward assessing judgment, EQ, and collaboration under uncertainty — because the technical baseline gets amplified by AI tools, the human differentiator moves elsewhere. Sam Kececi’s Sentience skips LeetCode interviews entirely, opting for hands-on working sessions instead. He hires for EQ plus IQ, and looks for candidates who care about impact, not just technology.

What’s the risk of over-automating team management? The main risk is losing the human trust layer — with employees, with customers, with investors. AI agents can handle routing and coordination, but they can’t replace the strategic decisions that determine a company’s direction, nor the interpersonal trust that keeps a team through hard times. Founders who over-automate risk making their companies feel transactional rather than human.

How do you know when an AI agent is ready to take on a task without human review? Currently, most production-quality personal AI operates with human-in-the-loop guardrails on every action that has real-world consequences. The escalation pattern used at Sentience is: agents flag important moments back to the founder via notification, so a human makes the final call on anything that could damage a relationship or misallocate resources.

Will AI agents eventually replace founders too? Not in the strategic-decision or vision-setting sense. Founders still hold the taste, the values, and the market judgment that determine what the company becomes. What AI agents replace is the operational drag that used to consume founder attention — the switchboard work — freeing founders to focus on the strategic layer where they can’t be replaced.

How does managing agents scale differently than managing humans? Theoretically, agent management scales infinitely — you can spin up 10 or 100 agents for the same marginal cost. Human management scales sub-linearly because of coordination overhead. In practice, the bottleneck shifts to the founder’s ability to review and coordinate agent output. The new limit isn’t headcount; it’s judgment throughput.

Where can I learn more about how Sentience actually runs its team with agents? Sam Kececi walks through the specifics on the TwoSetAI podcast, including how their engineering standup runs entirely through Sentience with agents that sometimes disagree with each other or crack jokes. Watch the full episode on the TwoSetAI YouTube channel.