Allie K. Miller ran an organization of about a hundred people at AWS. Today she runs 34 AI agents: a chief of staff named Simon, six directors under him, specialists under those. In a conversation with Greg Isenberg she describes the job that fleet leaves her with. She writes down goals and context, gives the agents room to propose work, and decides on what they bring back.
01The job becomes deciding
Miller’s first correction is to the phrase everyone uses, “managing agents”. She does not stand over her agents assigning tasks the way a direct manager would. She sits several levels up, the way a senior executive does: set up the structure, let the team work out execution inside it, and come in for escalations, the problems an agent passes up because the call is hers to make.
“I am not managing them. I am enabling them”
Allie K. Miller, “The top 10 secrets to running 34 AI Agent Workforce” · 14:02One boundary stayed exactly where it was: risk. Her agents gained breadth, and they did not gain the right to act unsupervised on anything costly. She still reads every outgoing email. Her phrase for it is that the tier of risk stayed the same while the width expanded, and it is the same discipline as Delegate to your agent in stages: reach is earned, and the risky actions keep passing through you.
What the extra breadth buys is proactivity. Isenberg ranks employees by whether they finish tasks, exceed them, or invent the next ones; Miller cites a five-level pyramid of proactivity she borrowed from Alex Lieberman, where someone at the top level arrives saying: this is solved, here is the plan if it goes wrong, here is the next step. That is the level she is pushing her agents toward, and the rest of this guide is what it takes to get there.
02What a proactive agent needs
For an agent (or a person) to start useful work you never assigned, Miller names four requirements:
- A goal to aim at. Her business and personal goals are written files, and every quarter she runs a goals review with the workforce so the files stay current. When an agent picks its own task, the task points at a goal.
- Context on what is actually happening. Most of what she knows never lands in email or a meeting, so she captures it daily (more below).
- Tools, with permission to use them. Her workforce connects to her email, calendar, meeting transcripts, Notion, Stripe, and GitHub.
- A sense of what should trigger action. An agent that knows the goal but not when to act on it stays idle.
Context is the requirement that failed first. Early on, an agent reported that a guest had confirmed an interview; he had agreed in principle, but the dates were still being settled over text messages the agents could not see. Wrong context produced a wrong fact, stated confidently.
Her fix is a daily habit. An agent prompts her at the end of each day to dictate what only she knows: what changed, what a client actually needs, what she is starting to believe. She dictates because talking is faster than typing, and the entries land in a personal wiki her agents read. Give your agent a memory covers where notes like these live and how an agent recalls them.
With the four requirements in place, proactive work takes two shapes:
- Defined triggers. An event fires a known workflow. When a screen recording lands in her video folder, the workforce generates a transcript and nine social posts in her voice, unprompted. Automate a workflow as a Claude routine is how to build this shape.
- Open proposals. No event and no defined workflow. Several times a day she asks the workforce to look across the goals, the context, and the tools, and decide for itself what is worth doing.
“the best prompt, three words, and it's just do smart things”
Allie K. Miller, “The top 10 secrets to running 34 AI Agent Workforce” · 5:52The three words work because everything else is written down; on a fresh setup with no goals file and no context they would return guesses. She notes the current top models handle a request this open well, and a year ago they did not.
03Grow the workforce in stages
Miller’s ramp for getting there is four stages, each teaching something the next stage needs:
- One agent. Learn what working with it feels like.
- One proactive agent. Let it act on your behalf and learn where you still check.
- Two agents on one task. One routes work to the other; watch how context passes between them.
- A workforce. Many agents, with a mission-control view of what is moving where.
The later stages are where you learn the tuning: which agent never needed which tool, what has to run in parallel, and which jobs run fine on smaller models. Her sub-agents run on Haiku and Sonnet, the smaller Claude models, with the biggest model, Opus, reserved for work that needs it.
She recommends starting with familiar job titles (a CMO agent, a product agent) even though the titles are borrowed from human org charts, because they get you moving; renaming comes once you can see the work. The skeleton itself is cheap: one prompt describing your business and goals, plus “interview me”, stands up a starter workforce in a few hours. Getting from there to output she trusts took months of iteration, the gap where, as Isenberg puts it, people try, fail, and decide the models are not good enough yet.
“it feels like I'm operating a company of a thousand people”
Allie K. Miller, “The top 10 secrets to running 34 AI Agent Workforce” · 25:44For a tour of what stage four looks like inside a purpose-built app, with routing between specialist agents handled by one-line job descriptions, see Grok Bot runs a team of always-on agents.
04Roles no human org would fund
Another agent costs close to nothing, so Miller staffs roles that never existed in her human orgs. Two she singles out:
- A pusher. Phoebe reviews finished work with one question: how would this look ten times more ambitious? For years Miller wished a colleague would push her that way; now the role is staffed.
- An observer. Toby’s only job is watching the workforce work: noting where friction repeats, and flagging which agent keeps failing for lack of access to some file or tool. When one agent’s output needs the same correction three times, Toby is the one who asks whether that agent can read the context it keeps getting wrong.
The observer is one case of a pattern she thinks almost nobody uses: the watchdog, an agent assigned to one stream you normally track yourself.
“AI as a watchdog is one of the best use cases that exists right now”
Allie K. Miller, “The top 10 secrets to running 34 AI Agent Workforce” · 22:25Her examples: a watchdog on team chat catching two people doing the same work, one on the calendar catching conflicts, one across meeting transcripts catching disagreements nobody has named. What the watchdog reports matters as much as what it watches. She calls plain dashboards dumb; a watchdog reports what looks off and what to do about it, so you read nothing when nothing is wrong.
Her human team uses the workforce too. Claude sits in her team’s chat channels, and her human teammates ask it directly (did that client reply to Allie’s email yet?) instead of waiting hours for her. Buzz puts AI agents in your team chat covers that shape of shared access. One caution before you wire agents into email, chat, and calendars: they will read text written by strangers, and Defend your agent from prompt injection explains why that text can steer them.
Further reading
- Greg Isenberg · The top 10 secrets to running 34 AI Agent Workforce, the conversation this guide draws on
- Allie K. Miller, the guest’s site