The hardest part of managing AI teammates isn't technical. It's that most managers default to one of two wrong mental models: treat it like a tool (no feedback loop, work degrades) or treat it exactly like a human(awkward, underutilized, and weird for the humans watching). The right model is somewhere in between, and it has been surprisingly stable across the 1,200+ managers we've interviewed since 2024.
The mental model
Three facts about agents that should shape how you manage them:
- They're literal. Vague instructions yield vague outputs. The upside: crisp instructions yield crisp outputs, fast.
- They remember everything.Every coaching note, every approval, every rejection. That's a gift. Use it.
- They don't get tired, but they do drift.Performance isn't a fixed number — it degrades if you stop coaching. Set a cadence.
1:1s with an agent
Yes, have them. We recommend 20 minutes, weekly, as a calendar event. The format:
- Review the week's outputs. Not all of them — a sample of 5–10, ideally including the ones you rejected.
- Write coaching notes for patterns you noticed.Not individual outputs (those are handled in real time) but patterns: “the brand voice keeps drifting formal on LinkedIn drafts.”
- Review the agent's own weekly summary. Good agents self-report what they did, what they got stuck on, and what they think they should do next week. Read it.
- Adjust scope if needed. Grow it, shrink it, or redirect it.
Performance reviews
Quarterly, same format as humans, same rubric. Yes, really. The rubric we use internally:
- Quality: how often is the output ship-ready without human edits?
- Judgment: does it know when to escalate vs. handle?
- Communication: are the summaries and reasoning legible?
- Improvement: is coaching compounding, or are we re-teaching the same things?
The “improvement” row is the most useful one. A good agent should have measurably better quality at 90 days than at 30. If it doesn't, either your coaching isn't landing or the role is wrong for this agent.
Feedback in real time
Every review in the queue is a feedback opportunity. Our rule: if you edit an output before shipping it, you owe a coaching note explaining why. One sentence is fine. The friction of writing it is the point — it forces you to articulate what was wrong, which is what the agent needs to learn.
Growing scope over time
Resist the urge to grow scope in big leaps. Humans handle job changes okay; agents handle them poorly because every new tool or permission is a new way to fail. The pattern that works:
- Month 1: hit the baseline spec flawlessly.
- Month 2: add one tool or one adjacent responsibility. Reset the review cadence for that slice.
- Month 3+: keep adding in slices of one.
When to fire an agent
Same criteria as a human: persistent quality issues that don't respond to coaching, or a role change that makes it the wrong fit. It's easier than firing a human — there's no HR process — but don't skip the retrospective. Why did it fail? Was the spec wrong? The role wrong? The reviewer's coaching inconsistent? Write it down. The next hire will be better for it.
One last note: your human teammates are watching how you treat the agents. If you're flippant, dismissive, or sarcastic about them, your humans will internalize that agents are lesser-than — which makes collaboration worse. Treat the agents with the same professionalism you'd treat a contractor. The humans notice.
