# Moving up the stack

A mental model I’ve found useful for thinking about how agents are changing work:

Most organizations have three(+) layers.

• →Individual contributors (ICs) do the work.
• →Managers do a little of the work themselves, hand out direction, and check what the ICs produce.
• →Directors/VPs/execs keep the bigger picture in their heads. They set priorities, sequence the work, and communicate across teams. They might review the actual work occasionally, but rarely. Their managers largely handle that for them.

Directors are wide but shallow, ICs are narrow but deep, and managers sit somewhere in between. The pattern holds across functions, whether in engineering, product, marketing, or anywhere else.

## Now

Agents are shifting people up the stack.

By adding another 5–10 people (agents) each IC can direct, we’ve turned every IC into a manager. And consequently, we’ve turned every manager into a director.

## Next

I think we’re heading one more notch up. As agents get more capable, they’ll take over more of the manager’s work too, and we’ll all move toward director-level work.

In that world, reviewing diffs or individual line changes is extremely rare. Even reading the output of individual sessions matters much less. You might still want to sometimes, but usually you won’t. We’re already seeing early signs of this on our team (when’s the last time you read a workflow transcript?), and I expect a lot of organizations are moving the same way.

## What directors actually do

Traditionally, a director might have five teams and forty people under them, with hundreds of things in flight. Their work looks very different from an IC’s. It’s mostly:

• →Compressing information into a big picture
• →Unblocking when issues come up
• →Setting priorities and sequencing work
• →Dispatching changes in direction or priority
• →Coordinating with other directors across team boundaries

All of it depends on the same thing: a clear picture of the whole, without drowning in the parts.

## What that could mean for our tools

If people are doing director-shaped work, they will need director-shaped tools. In tools like Claude and Codex, that could translate to:

• →Summaries and dashboards first, with detail you can drill into as needed
• →Regular status updates pushed to you, so you don’t have to go looking
• →Non-actionable information tucked away until you need to dig in
• →A queue of judgment calls, where blockers, escalations and questions surface on their own
• →The ability to steer with a message: one instruction to a whole “team,” which delegates it to the agents within it

Lots TBD with how that interface could look & work. But the director’s workflow could be a good pattern to keep in mind as we think about what tools we want to build.