canonical_url:
Two Fortune 500 execs stood on a summit stage last week and gave opposite answers to one question: is an AI agent a colleague or a tool?
One names his agents and seats them in reviews. The other won't call them colleagues. I watched the debate go by and realized I'd stopped caring about the noun a while ago, because it never once changed what I did with my backlog on a Monday.
So here's the thing I actually changed, and the part of it that broke.
The old default: sort by who's free
For years my planning loop was the same. Look at what needs doing. Look at who has capacity. Assign. The unit was always the person. "Who's free" was the first question.
When agents showed up in the loop, I just slotted them in as another row in the capacity table. Same question, one more name. That's the "tool" answer in practice - an agent is a faster hand, you point it at whatever's next.
It worked until it didn't. The agent would happily take a task that needed a human to even scope it correctly, produce something plausible, and I'd find out two steps later that the plausible thing was wrong in a way that cost me a half-day to unwind.
The change: sort by shape first
I flipped the order. Before I look at who's free, I sort the backlog by shape.
Two buckets. Person-shaped work needs judgment, lives in ambiguity, depends on taste or a relationship I can't write down. Agent-shaped work is defined, repeatable, and gate-able - I can describe the input, the output, and the check it has to pass.
The discipline is that I write the agent's slice like a contractor's statement of work, not a chat prompt:
# agent slice — scoped like an SOW, not a vibe
task: regenerate API client from updated OpenAPI spec
input:
- openapi.yaml (v3.1, committed)
- existing client at src/clients/
output:
- regenerated client, same public surface
gate:
- all existing contract tests pass
- public exports diff reviewed by a human before merge
owner_of_outcome: me # not the agent
If I can't fill in gate and output cleanly, that's my signal: this isn't agent-shaped yet. It's person-shaped work I was about to mislabel because I wanted it off my plate.
I ran this across about 40 backlog items. Roughly a third of what I'd have handed to an agent under the old "who's free" sort failed the shape test and went back to a human.
What broke
Two things broke, and both were useful.
First, my sense of which work was "important." I'd been guarding a pile of tasks as too critical to automate. Half of them were just defined work I was emotionally attached to. They sorted cleanly into the agent bucket and ran fine. The status I'd assigned them was about me, not the work.
Second - and this is the one I underscoped - the gate is only as good as the human watching it. There's research showing that once you treat an agent like a teammate, you review its output less carefully. Naming the thing lowers your guard. So the "colleague" framing isn't warmth, it's a quiet accountability leak. I caught myself rubber-stamping a passing gate because the agent had "been reliable lately." A tool doesn't earn trust. Work does, run by run.
There's an old line going around that fits: you automate the boring, and then humans just manage the failure modes. That's not a downgrade of the human job. The failure modes ARE the job now. Managing the part the agent can't hold is the work that's left, and it's the work that compounds.
Why this is a career thing, not just a workflow thing
The same exec who named his agents also said the hardest part isn't the technology, it's the managers. I think that's exactly right, and it cuts both ways.
If the manager is the bottleneck, the manager is also the leverage. The skill that pays off from here isn't prompt-writing - that floor keeps dropping. It's the older skill of looking at a pile of work and knowing fast which parts are person-shaped and which you can scope, gate, and let run. That's "golden age of the generalist" energy: when the specialization tax drops, the person who holds judgment AND can design the work is the one who wins.
The execs on stage were missing a practitioner, so they argued about the label. The label was never the leadership problem. The work was.
So I'm curious where you've landed: when an agent is in your loop, do you still plan by who's free, or have you started cutting the work by shape first? And where has the gate let something through on you?
Top comments (1)
honestly the shape sort breaks down on work that's half-and-half - the scoping is person-shaped but the execution is agent-shaped, and i keep wanting one bucket when it really needs splitting mid-task. i hand-waved that case in the article.