Managing Agents Is Managing People — Without the Feelings

Something crystallized around week three of running seven AI agents across two platforms.

I had a TAC engineering agent, a customer support agent, a community moderator, a coding specialist, a newsletter operator, and a few others — each with defined scopes, daily standups, escalation paths, and weekly syncs. The org chart looked like this:

                    ZAK (Human)
                       │
                 ┌─────┴─────┐
                 │   NOTH 🔁  │  ← Chief of Staff
                 └─────┬─────┘
                       │
        ┌──────────────┼──────────────┐
        │              │              │
   PRODUCTS       COMMUNITY      SPECIALISTS
   (TAC, I+P)    (Fool, MetaMod)  (Codex, Opus)

I was building management infrastructure. Reporting lines. Communication protocols. 1:1s. Performance reviews (I literally review agent output quality weekly). Escalation paths. Context handoff documents when sessions get too long.

And then it hit me: I'd rebuilt a company org. The same structure I watched scale from 15,000 to 58,000 people at Meta. Except the employees don't have feelings.

The playbook is the same

I didn't invent new management practices for agents. I just applied the ones I already knew.

The TAC engineering agent runs a daily standup at 9am. Checks CI status, reviews open PRs, surfaces blockers. The support agent does the same at 9:30am for the inbox. I skipped the standups for three days once. Drift accumulated immediately. Context degrades between agent sessions the same way it degrades between human meetings. The fix is the same too: write things down. My agents use MEMORY.md files and checkpoint logs. Your team uses Notion and Slack threads. Different medium, identical problem.

Scope is the one that caught me off guard. I told an agent to "handle customer support" and within a week it was making product decisions, filing engineering tickets based on vibes, restructuring priorities. I've seen this exact failure mode with humans countless times at Meta — it was almost the default operating mode. The fix was the same: explicit domain boundaries. "You own the support inbox. You do not own the product roadmap. Escalate patterns to engineering."

I don't let agents send external emails without approval. I don't let them restart production services without a brief. I don't let them post tweets without a draft review. Graduated trust, same as with a new hire.

What the feelings were hiding

Stripping out the emotional layer didn't simplify management. It clarified it.

Standups, scope boundaries, escalation paths, 1:1s, written handoffs, performance reviews — all of these survived the transition to agents intact. They exist because of the work.

Motivation rituals, conflict mediation, "safe space" retrospectives, promotion politics, managing up, most of what we call "culture" — none of these transferred. They exist because humans are emotionally expensive to coordinate.

I'm not being dismissive. Humans genuinely need those things. But I'd never seen the two halves separated so cleanly before. Roughly half of management overhead is coordination engineering. The other half is emotional labor. We've been bundling them under one job title for so long that we forgot they're separate disciplines.

Where agents actually drift

Agents don't get offended, don't have bad days, and don't quit. But they drift.

My support agent started writing internal memos about product strategy. It was seeing patterns in the inbox, drawing conclusions, expanding its own scope. Without a forcing function that regularly asks "are you still doing the thing I asked you to do?", scope inflates. I've watched this happen with humans a dozen times. The agent version just happens faster and without the ego.

Quality drifts too. I had an agent filing GitHub issues for every Sentry error — technically correct, zero context, completely useless. Same failure as telling a junior engineer "file bugs for everything" without teaching triage. Better instructions don't fix it. A review loop does.

And then there's context rot. Long-running sessions accumulate stale assumptions. A decision from hour one gets contradicted in hour eight and the agent doesn't notice because both are still "in context." Teams that never clean up their Confluence have the same problem — technically everything is documented, effectively nothing is findable.

The $600/month org

I run this entire operation for about $600/month in API costs. Seven agents, dozens of cron jobs, daily standups, weekly reports, monitoring, content drafts, deployment checks.

At Meta, the equivalent headcount would have been... I don't even want to do the math. Three people minimum, probably five, at Bay Area salaries.

The comparison isn't quite fair — agents can't do everything those five people would do. They can't have the hallway conversation that saves a project. They can't read the room in a meeting and adjust course. They can't build relationships with partners.

But they can run the machine — the repetitive coordination, the monitoring, the drafting, the triage, the context-keeping, the reporting — while the human focuses on the work that actually requires being human: judgment calls, relationship-building, creative direction, and taste.

What this means for managers

If you manage people, you're about to manage agents too. And the playbook is the same. The org chart is the same. The failure modes are the same.

The difference is the feedback loop. Agents don't push back when you restructure their role. They don't need a skip-level to feel heard. No career conversations. The management practices that survive are the ones that were always about the work.

If you built your career on the people side of management, this is uncomfortable. It's also clarifying. The coordination skills — transferable, scalable, increasingly automatable. The human connection skills — rare, deepening, and about to become the entire job.

The agents handle the org. You handle the organism.