People who orchestrate workers are managers.
People who orchestrate software are engineers.
People who orchestrate general intelligence inside software are… not talked about enough.
Welcome to the age of Agent Management.
It's become a cliché these days: brilliant engineers promoted into management roles often discover that they hate their new jobs. Engineers build systems governed by predictable logic. Managers build systems based on incentives, performance reviews, etc.
Imagine going from writing a unit test to conducting a 360 review—two different worlds.
When software breaks, you debug—step through the logic, isolate the bug, patch the issue. When a person you manage makes a mistake, you can't attach a debugger to their brain.
Human systems don’t compile neatly; they operate on uncertainty, incentives, and instructions that are rarely as clear-cut as code. And that’s exactly why great management is hard.
Effective management requires clear goals, aligned objectives, precise communication, IQ/EQ balance, and thoughtful performance evaluation. And the complexity of management only increases depending on whom (or what) you're managing:
Determining if Worker #17 hit their production target is simple; evaluating an executive’s overall impact and effectiveness is far more nuanced.
Now imagine scaling this complexity even further—beyond just human teams—to entirely new forms of intelligence.
AI isn’t human—and it’s also not an if statement. It’s an alien intelligence that can do things we associate with human reasoning, yet can also scale and replicate like software.
Managing these non-human intelligences draws on both engineering’s demand for precision and reliability and management’s art of coordination through goals and incentives, but it also calls for a new playbook altogether. The best Agent Managers will synthesize lessons from both disciplines, build an intuition for this alien intelligence, and try lots of new things—just as managers have iterated through approaches over centuries.
Agent Managers may be amazing engineers and amazing managers, but some may be neither. Some of the best Agent Managers of tomorrow might not even know how to code today.
Traditionally, managing ML models meant training them for clear-cut tasks: straightforward inputs, straightforward evaluations. Today’s ML models function more like knowledge workers—handling a more general class of task with autonomy, making success harder to measure.
Soon, applied ML will resemble managing executives who oversee specialized teams. To excel in this environment, you'll need to:
We’re moving from getting a model to do a task, to managing a team of models to run a department.
Evaluating autonomous systems is inherently challenging. Just as assessing executives is an ongoing topic of debate, evaluating advanced agentic systems demands increasingly sophisticated frameworks. As models grow more capable → we grant more autonomy → they perform more valuable tasks → evaluating and guaranteeing performance become increasingly complex.
This cycle will continue. Mastering evals is still underrated.
Managing alien intelligence only works if the “aliens” themselves are up to the task. You can’t transform a mediocre performer into a star just by giving more pep talks. We need better “recruits.”
That’s where our applied ML research comes in. For your agentic system to be effective, you need the models themselves to perform. And unlike humans, ML models actually do allow researchers to peer into their "brains" and make changes (fine-tune, post-train, etc.) so that the models can work together to form the best agentic system. ML research and Agent Management will work in tandem to push the frontier of what's possible.
We’re looking for people who understand this evolution and want to shape this moment. You’ll thrive here if you: