# My top secrets to running an AI Agent Workforce

Greg Isenberg interviews Allie K. Miller on shifting from *managing* AI agents to *enabling* them: a 34-agent org led by chief of staff Simon, a three-word prompt, a daily AI diary, watchdogs, and building the software factory before the product.

Tags: youtube-inbox, source, agents, workforce, allie-k-miller, greg-isenberg, claude

## Summary

Allie runs 34 AI agents under Simon (AI chief of staff) and six Friends-named directors. She sits three rungs above them, keeps the risk tier (she still checks outbound email) while expanding breadth, and several times a day prompts the workforce with "do smart things" on top of full business context. Greg and Allie then argue SaaS is not collapsing on enterprise timelines, while consumer software rewards taste and distribution.

## Key points

- **Mindset**: "managing agents" is early-2026 talk. Allie frames herself as SVP three rungs up — infrastructure, escalations, liability/final say — not a direct manager assigning tasks. People come to AI partly to *escape* management work; keep the empowerment, drop the admin.
- **Org**: 34 agents. Simon = AI chief of staff. Six directors (education, client work, ops, marketing, product, plus **Phoebe** as chief dreaming officer / 10x layer). **Toby** watches the workforce and logs friction and access gaps. Start with 2015 titles, then hire roles a human budget would never fund.
- **Three-word prompt**: "do smart things." Context includes business/personal goals, meeting transcripts, email, calendar, Notion, Stripe, Supabase, GitHub, friends/family docs. Frontier models (Fable 5 / GPT-5.6 class, per talk) can now take that vague prompt; a year ago they could not. Risk tier unchanged (still reviews emails); width expands.
- **Pyramid of proactivity**: Greg's three employee types (doesn't finish / finishes assigned / invents new work) mapped to Alex Lieberman's five levels. Level 5 solves it, plans failure, and next steps. Quarterly goals review with the workforce so net-new tasks stay goal-oriented; treat agents as product leads, not only engineers.
- **Proactive 2026**: 2025/H1 2026 was trigger automations (drop a screen recording → transcript + nine social posts). H2 2026 target is *undefined* probabilistic workflows: goal + tools + permission + a sense of what would trigger action, without Allie as first domino.
- **Queryable company / daily diary**: meetings/email/Slack miss a lot. A daily prompt asks her to **dictate** (claimed 4× faster than typing) uncodified context into a personal wiki (~80–86 entries). Early failure mode: agents assumed calendar/email truth while dates lived in iMessage after an MCP break.
- **How to start**: one agent → one *proactive* agent → two agents collaborating/routing → a workforce with mission control. One public prompt: describe the business and goals, ask the model to interview you, target saving hours/capping meetings. Baseline + tools in **under three hours**; 90%+ is iteration. Sub-agents on Haiku/Sonnet; reserve Opus for heavy work.
- **Multiplayer**: Slack channel **Loop Alley** — humans talk to the AI workforce (e.g. "did the client reply to Allie's email?"). Claude in chat channels; story of Claude salute-emoji-reacting to a late-night rant.
- **AI as watchdog** (underused): duplicate work in Slack, calendar conflicts, meeting disagreements, contract compare-and-contrast. Dashboards are "dumb"; want anomaly detection plus insights-to-action (what to post tomorrow, write the script).
- **Factory then product**: for the public **AI First Index** (Fortune 500 exec interviews, ~16 dimensions), they built primitives (login, payments, social sharing, newsletters) instead of one-shotting the app. First product already profitable; next products ship faster. "Dark headless factory" / measure twice (foundations) cut once.
- **SaaS / enterprise**: mediocre software dies on a *years* timeline, not overnight — most paid AI users still are not on Codex/Claude Code; enterprises are bandwidth-constrained and will not rebuild DocuSign/Salesforce until make+maintain+secure is ~95% easy and under a few hours. Buyers want a vendor to call, someone to blame, and not to lag AI-lab partner day-0 access.
- **Consumer**: science → art. Taste, video, word-of-mouth, B2B influencer trust are high-value bottlenecks; local HTML → iOS still messy. YC mix has skewed B2B; Allie cites a no-code face-model photo app with ~300k users as B2C proof. Agent-first versions of charting consumer apps (Century Tower mentioned) as a search avenue.
- **Research tricks**: YC Instagram asks (agent-first software; assume 18 months out); Matt Van Horn **last30days** GitHub skill (fan-out news synthesis into a Claude wiki); listen for CMO "fear zones" (agent discovery, agent-first shopping, brand consideration).

## Quotes / memorable lines

- "I feel like the term managing agents is wrong."
- "The best prompt… three words… do smart things."
- "The tier of risk has stayed the same but the width has expanded."
- "I want my whole company to be queryable."
- "If I give all of them job titles that existed in 2015… I am operating in 2015 org structure."
- "Stop relying on only yourself to find these blockers. AI as a watchdog is one of the best use cases that exists right now."
- "Instead of building out the thing, build the factory for the thing."
- "You also want someone to blame."
- "Look for the bottlenecks then evaluate the value of fixing those bottlenecks and then pick the bottleneck that is high value to fix."
- "Leaning into the weirdness of what it looks like to have not just an AI workforce, but to have a multiplayer AI workforce."

## Sources

- YouTube: https://www.youtube.com/watch?v=EzQAgnjTq2k
- Raw ingest: `raw/youtube-inbox/EzQAgnjTq2k/` (metadata.json, transcript.txt, notes.md, source.url)
- Channel: Greg Isenberg https://www.youtube.com/channel/UCPjNBjflYl0-HQtUvOx0Ibw
- Guest: Allie K. Miller https://x.com/alliekmiller
- Transcript: Supadata v1 mode=native text=true, lang=en, 45542 chars. ASR spellings are not authoritative.

## Related pages

- [[hermes-agent]] (self-hosted agent org vs Allie's Claude/Codex workforce)
- [[youtube-hermes-kanban-multi-agent]] (durable multi-agent coordination + human gate)
- [[grok-bot]] (managed computer-use teammate vs a 34-agent org chart)
- [[grok-bot-vs-hermes-agent]]
- [[youtube-hermes-desktop-plugins]]