The Best AI Agent Managers Were Already Managing Humans

Everyone is scrambling to hire prompt engineers. I keep watching companies miss the obvious.

The people on my teams who get the most out of agents are the ones who have spent years telling humans what to do, holding them accountable, and knowing when to cut someone loose. Prompting is a management skill. It has always been a management skill. The market just refuses to see it that way.

Now I have company. Ivan Burazin, cofounder and CEO of Daytona, is running 80 agents across a 30-person team. His 16 engineers each run an average of five agents. Nobody at the company writes code anymore. And the thing he keeps noticing? Staff who have managed other people are dramatically better at prompting agents than those who have not.

"We've found that people who manage people and understand how to communicate what they want — inputs versus outputs — they're much better at running agents," Burazin told Business Insider.

Yeah. I know.

Delegation Is Delegation

Think about what good management actually requires. You have to articulate what you want clearly enough that someone else can execute it without you standing over their shoulder. You have to define success upfront. You have to set guardrails without micromanaging every step. You have to know when the output is wrong and why, and course-correct without destroying the working relationship.

That is exactly what prompting an agent requires.

When I am writing a system prompt for an agent in Cadderly, I am not writing code. I am writing a job description, an operating procedure, and a performance rubric all at once. I am telling the agent what it owns, what it does not touch, how to escalate, and what good looks like. Any manager who has ever written an SOC or onboarded a new hire has done this exact work. They just did it with humans.

The engineers who struggle with agents are often the ones who want to control every step of execution. They write prompts like they are writing scripts. They over-specify the how and under-specify the what. That is a shit recipe for success with humans, and it is a shit recipe for success with agents.

The Talent Market Is Looking in the Wrong Place

Companies are posting job listings for "Prompt Engineers" and filtering for people with ML backgrounds or Python chops. Meanwhile, the person who spent five years managing a 12-person customer ops team, who knows how to write a clear brief and hold people accountable to outcomes, is sitting right there being overlooked.

According to Deloitte, 74% of organizations expect to be using agentic AI at least moderately within two years. Only 21% currently have a mature model for agent governance in place. That gap is not a technical problem. It is a management problem.

IDC forecasts that by 2026, 40% of G2000 job roles will involve direct interaction with AI systems. The people best equipped for that are the ones who already know how to run a team, set expectations, and own outcomes regardless of who or what produced the work.

Bernard Marr put it plainly: the priority is to clearly define goals, set guardrails, and establish criteria for automated decision-making and escalation. That is management. That has always been management.

What I See on My Own Teams

We have been building Cadderly as a coordination agent since before most people were talking about multi-agent systems. Intent recognition, MCP integration, A2A coordination. The hardest part has never been the architecture. The hardest part is always the instruction layer.

The people who write the best agent instructions on my teams share a few traits:

• They think in outcomes, not steps. They define what done looks like before they define how to get there.
• They anticipate failure modes. They ask "what will this agent do when it hits an edge case" the same way a manager asks "what does my rep do when a deal goes sideways."
• They iterate fast. They treat a bad output like a performance conversation, not a catastrophe. They adjust and move on.
• They know when to escalate. They build in checkpoints where the agent hands off to a human, because they understand that autonomy has limits.

None of those traits come from knowing Python. They come from managing people.

The Prompt Is the Job Description

Here is the mental model I use and share with everyone I work with.

Your system prompt is the job description. Your context window is the onboarding document. Your few-shot examples are the training sessions. Your evals are the performance reviews. Your escalation paths are the org chart.

If you have built a team before, you already know how to do all of this. You just have to stop thinking of the agent as software and start thinking of it as a new hire with a very specific skill set, infinite patience, zero ego, and a tendency to hallucinate when under-briefed.

That last part matters. A misconfigured instruction does not produce one wrong output, it produces a chain of wrong outputs. Anyone who has ever given ambiguous direction to a team and watched the downstream damage unfold over two weeks knows exactly what that feels like.

This Changes Who You Should Be Hiring

If you are building an agentic product or deploying agents internally, rethink your hiring criteria.

Stop filtering exclusively for technical backgrounds. Start looking for:

• People who have managed cross-functional teams
• People who write clear, outcome-oriented briefs
• People who have run operations and built SOPs
• People who know how to hold others accountable without over-controlling execution
• People who have fired someone. Seriously. That judgment is the same judgment you need to know when an agent is failing and needs to be retooled or replaced.

Do your job or I will replace you. That applies to agents too. The people who understand that instinctively are your best agent managers.

Daytona raised a $48 million Series B. Burazin is not running a side experiment. He is running a real company on 80 agents with 30 people, and the management insight he is surfacing is the same one I have been watching play out in every agentic deployment I have touched.

The unlock is already sitting in your org. You probably just have not looked at them yet.

---

If you are thinking through how to structure your agent teams, what to look for in the people running them, or how coordination architecture like Cadderly fits into the picture, reach out. I am happy to talk through what we are seeing in the field.