# The $5K-a-month AI-agent agency just became a public playbook

URL: https://www.thedeepfeed.ai/posts/2026-05-12-greg-isenberg-managed-agent-business-playbook/
Category: Business
Published: 2026-04-22
Author: the-deep-feed
Tags: ai-agents, agent-agency, hermes-agent, orgo, composio, solopreneur
Kind: deep

> Greg Isenberg's May 12 podcast with Nick Vasiles of Orgo took the underground AI-agent-agency operator pattern and turned it into a 47-minute YouTube playbook. The stack, the unit economics, and the daylight problem are now public.

## TL;DR

- Greg Isenberg's May 12 podcast with Nick Vasiles is the **first public, end-to-end productization** of the AI-agent-agency playbook: charge **$5K/month for OpenClaw-based agents, $10K/month for Hermes-based agents**, sell into legacy verticals (marketing, law, insurance, manufacturing, wholesale, real estate). The stack is fully named: **Hermes Agent + Orgo + Composio + AgentMail + Obsidian + Codex/Claude Code + GPT-5.5**.
- The economic novelty is the **anti-friction offer**: unlimited agents, unlimited usage, unlimited monitoring, security included. Nick concedes in the transcript that customers think they need 10–100 agents but the real number is **one to three**. The unlimited framing is **psychology, not capacity** — it removes credit-counting from the conversation.
- **The harness pricing tier is real and it's already segmenting.** [OpenClaw](https://github.com/openclaw/openclaw) at 371K stars is *"commoditized already"* (Nick's word) at $5K/month; [Hermes Agent](https://github.com/NousResearch/hermes-agent) at 147K stars commands $10K/month because it ships continuous self-evolving capability. Same agent shape, 2× price, on harness brand alone.
- **Greg is running the same play three times in a week.** Within the May 9–12 window he shipped: (1) ["Agent franchise kits"](https://x.com/gregisenberg/status/2052988834928865000), (2) ["Outcome bounties"](https://x.com/gregisenberg/status/2052991022358593000) plus a ["scorecard for rating AI agent businesses"](https://x.com/gregisenberg/status/2052991022358593000), and (3) this 47-minute Nick podcast. The thesis under all three: **agent-agency-as-a-service is the 2026 drop-shipping equivalent** — a high-leverage solopreneur model gated only on AI-fluency, not capital or headcount.
- **The daylight problem is the bottleneck.** Underground operators were already at $5K MRR per customer in private discord communities. Publishing the playbook collapses information asymmetry, compresses margin, and **brings the price floor with it** — exactly what happened to Shopify drop-shipping in 2018-2019. Within 12 months the floor moves to $1,500-$2,500/month, the differentiator becomes vertical-specific skills + harness-engineering depth, and the agencies that didn't build *fulfillment moats* (watchdogs, observability, sub-niche expertise) get displaced by content-marketed lookalikes.

On May 12, 2026, at 6:06 PM UTC, [Greg Isenberg](https://x.com/gregisenberg) (host of the Startup Ideas Podcast, CEO of Late Checkout, and one of the most-listened-to voices on solopreneur product strategy on X) posted a [10-point thread](https://x.com/gregisenberg/status/2054261832718889216) summarizing his new 47-minute YouTube episode with [Nick Vasiles](https://x.com/nickvasiles) of [Orgo](https://orgo.ai/). The episode's title is the entire pitch: [*The $1M+ Solo AI Agent Business (Full Course)*](https://youtu.be/BI-MNjm1tTQ). It opens with Greg saying, on camera:

> *"People are charging $5,000 a month per customer to build and manage agents for them. This is a startup idea I wish more people would do."*

The thread is 678 views, 25 likes, 25 bookmarks at first measurement. The video racked up **1,431 views and 140 likes in its first 21 hours**, modest reach by Greg's standards (his organic ceiling is 200K+ views in 24 hours). But the **shape of the audience matters more than the count**: this is the cohort of operators who already build with AI agents and are looking for a productization model. Greg's average bookmark-to-like ratio on this kind of post runs above 1.0, and the bookmark count here is climbing in lockstep with likes — the same signal that the [agent-harness-engineering essay](/posts/2026-05-09-agent-harness-engineering-the-discipline/) gave us three days earlier.

The substance of the episode is unusual for the genre. Most Greg podcasts are idea-of-the-week riffs in the *"here's a hole in a market, here's who could fill it, go build"* shape. This one is a **full operating manual** — every step named, every tool specified — for a business shape that has been running quietly in operator communities since at least Q4 2025: **a one-person agency that builds and manages AI agents for SMB customers at $5K to $10K per month per agent, with the operator setting up the stack inside cloud VMs, configuring the harness, wiring the integrations, and collecting recurring revenue indefinitely**.

This post reads the episode as it actually is. The stack the operators have converged on, with line-level citations. The unit economics the playbook is implicitly proposing and which it leaves out. The competitive shape that emerges the moment the playbook is no longer underground. And the question that the practitioners on X are already debating: whether this is the 2026 equivalent of agency drop-shipping (a real income stream that displaces incumbents and then commoditizes), or the 2026 equivalent of NFT minting tutorials (a content-marketed pyramid funnel where the playbook itself is the product).

![The agent-agency stack as an exploded diagram: Hermes/OpenClaw on Orgo computers, Composio auth, AgentMail, Obsidian, perimeter watchdogs.](/post-images/2026-05-12-greg-isenberg-managed-agent-business-playbook/hero-stack.jpg)

## The offer, in plain terms

The episode opens with what Nick frames as *"the offer"* — the productization wrapper that distinguishes a managed-agent business from selling consulting hours. The pitch is **abundance-coded**: unlimited agents, unlimited usage, unlimited monitoring, security included, ongoing changes included, all for a single recurring monthly fee. Nick's repeated number throughout the episode is **$5,000 per month per customer** for the standard tier; **$10,000 per month** for a Hermes-Agent-based tier where the harness is sold as the premium brand.

The "unlimited" framing is the part that needs unpacking. Nick is explicit in the transcript that it is not a capacity claim — it is a psychological frame:

> *"It's not that the customer is going to actually need unlimited agents. They're not going to need unlimited tokens, but they might they might think they do. In reality, they might think they need five agents, 10 agents, 100 agents, when really one, two, maybe three agents goes such a far way."*

The operator's actual exposure on token costs is therefore bounded by a ceiling the customer rarely approaches. The structure resembles the [unlimited-design-revisions subscription model](https://www.designjoy.co/) that became the canonical productized-service template after 2020. The math closes for the same reason the design-subscription math closes: most customers do not run anywhere close to their plan ceiling, and the operator can margin the rest. The risk vector is identical too: the **one customer who actually does** can burn the unit economics, which is why Nick spends roughly nine minutes of the episode on guardrails. The list is specific: watchdogs, observability, scope-limiting via Trello, "one to two requests at a time delivered in under 48 hours."

The other half of the offer is what Nick calls *"removing all the friction"*. The customer does not see tokens, models, credits, or infrastructure. They get what Greg verbally re-frames in the transcript as *"selling an AI employee, not selling an AI agent"*. Nick's recurring metaphor is **"digital employee"** — a phrase that the post-Sequoia [services-as-software thesis](/posts/2026-04-30-sequoia-services-as-software-thesis/) made canonical six weeks ago. The agent-agency model is, in effect, the **fractional, multi-tenant version** of that thesis: instead of a vertical SaaS company replacing services revenue for one industry, a solopreneur replaces it for one mid-market customer at a time, picking up $5K–$10K MRR on each one.

## The verticals Nick named, in his order

The transcript spends roughly six minutes on industry selection. Nick names two industries explicitly to avoid: **healthcare and finance**, citing *"high regulatory burden and red tape"* as the reason. He then names six verticals where the playbook is reported as working:

| Vertical | Why Nick picked it |
|---|---|
| **Marketing agencies** | Lots of operations work; agency owners are already AI-curious |
| **Law firms** | Large legacy industry; executive-level pain abstracts cleanly |
| **Insurance agencies** | People businesses; high inefficiency surface |
| **Manufacturers** | Want to be "AI native" but don't know how |
| **Wholesalers** | Same shape as manufacturing; ops-heavy |
| **Real estate agencies** | Demonstrably willing to pay for tooling |

The unifying claim Nick makes is that **every executive in every industry has the same problems**: too many emails, too many meetings, too many follow-ups, too many open loops. Solving the executive's day-one problems is the wedge. The vertical-specific skills get layered on top, after the agent has been adopted at the top of the org. This is the same pattern that [outbound-sales tooling companies](https://www.theinformation.com/articles/the-rise-of-the-go-to-market-agents) used to land in mid-market accounts: solve the obvious pain at the executive level, then expand horizontally.

Greg pushes back inside the episode with a *"diverge then converge"* framing that is worth quoting because it inverts the conventional niche-first advice:

> *"You don't have to start super niche from the beginning. In fact, you can always niche down after trying a marketing agency, trying a law firm, trying all these different industries, seeing what works well for you, where the market pulls you, and then going super vertical."*

The implicit operational claim is that **the marginal cost of trying a new vertical is low once the harness stack is set up** — because the harness is the same; only the skills and the Obsidian vault content change. This is consistent with how [Cline ships `use_subagents`](https://github.com/cline/cline) and how OpenClaw's skills directory works: the same agent runtime can adopt a new domain by adopting new context, not by being re-engineered.

![Six verticals (marketing, law, insurance, manufacturing, wholesale, real estate); healthcare and finance crossed out as too regulated.](/post-images/2026-05-12-greg-isenberg-managed-agent-business-playbook/verticals-grid.jpg)

## The named stack, with line-level citations

The episode's most useful section for an operator is the inventory of tools Nick says he is actually using. The transcript names twelve specific products, and for each one Nick gives a one-line justification. The full inventory:

| Layer | Tool | What it does | Nick's one-line reason |
|---|---|---|---|
| Meetings | [Granola](https://www.granola.ai/) | Meeting transcript + MCP | Syncs notes into Trello as agent requests |
| Customer board | [Trello](https://trello.com/) | Customer-facing kanban | Prevents scope creep; backlog/to-do/doing/done |
| Updates | Loom | Async video updates | "Send Loom updates at random hours" |
| Email | Superhuman | Inbox velocity | Keyboard-driven; bulk customer email |
| Internal PM | Asana | Internal-only board | Separate from the customer-facing Trello |
| Build agent | [Codex desktop](https://openai.com/codex/) | Configures customer agent | "Most generous, simplest, best desktop app" |
| Build agent (alt) | [Claude Code](https://www.anthropic.com/claude-code) | Configures customer agent | Used as fallback / co-builder |
| Customer agent | [Hermes Agent](https://github.com/NousResearch/hermes-agent) | Premium tier sold to customer | "Doesn't break, self-evolving" |
| Customer agent (alt) | [OpenClaw](https://github.com/openclaw/openclaw) | Standard tier sold to customer | "Commoditized already; $5K/month" |
| Cloud VM | [Orgo](https://orgo.ai/) | Sub-500ms cloud computer per agent | "We give your agent a computer to live in" |
| Auth + tools | [Composio](https://docs.composio.dev) | 1,000+ MCP toolkits + auth | "Biggest time sink solved" |
| Mailbox | [AgentMail](https://agentmail.to/) | Email account per agent | Mia the agent has her own inbox |
| Context | [Obsidian](https://obsidian.md/) | Knowledge base per customer | "Second brain"; markdown vault |
| Default model | GPT-5.5 | Tool-calling default | "So efficient with tool calls" |
| Light model | [GLM 5.1 (Z.AI)](https://z.ai/) | Cheap open-source fallback | "Best open-source option" |
| Long-horizon | Opus 4.7 (Anthropic) | Long-horizon coding handoff | Pair via Claude Code from inside Hermes |
| Live docs | Perplexity MCP | Up-to-date setup context | One of several MCP context sources |
| Live docs | Context7 MCP | Up-to-date GitHub docs | Cited specifically for Hermes setup |
| Live docs | Exa MCP | Real-time web search | Fan-out alongside Perplexity |
| Live docs | X MCP (Twitter) | Pull operator tweets as setup recipes | "I find amazing setups on Twitter" |

🔴 **The stack itself is the most important artifact of the episode.** It is the first time in a mainstream solopreneur podcast that the **operator-side tooling stack** has been named end-to-end without abstraction. That matters because each of the named products has a measurable economic incentive to confirm or deny what Nick says — Composio, AgentMail, Orgo, and Granola will all gain customers if his story is true. None of them have publicly contested the playbook in the 24 hours since it shipped; several have amplified it.

The model recommendation is also a quiet but consequential reversal. As recently as April 2026, the operator default was **Opus 4.7 inside Hermes Agent**, a setup Nick himself called *"a national security risk"* in [an April 18 tweet](https://x.com/nickvasiles/status/2048469000000000000) (597 likes). Four weeks later, the same operator's public recommendation is **GPT-5.5 as default**. The implicit story is that OpenAI's tool-calling generosity (paid plan usage limits, low token rates) has structurally outcompeted Anthropic's Opus tier on cost-per-completed-task — at least for the *managed-agent-for-SMB* workload shape. That is a non-trivial pricing-and-product win for OpenAI that the consumer benchmark coverage has not yet picked up.

## The harness-as-pricing-tier insight

The single most consequential paragraph in the entire transcript is roughly seven minutes in, when Nick gives the price stratification between harnesses verbatim:

> *"If you sell Hermes agents, you can charge 10K a month. Exactly. So, OpenClaw's commoditized already. You know, it's 5K a month."*

Nick goes further moments later when Greg asks him to name his stack with one-liner reasons each:

> *"Codex because it's more generous and it's simplest and they have the best desktop app. Hermes because it doesn't break and it's self-evolving. OpenClaw is not as self-evolving."*

🔴 **This is the moment a harness becomes a pricing tier.** Same agent shape (computer-using agent in a sandbox), same model (GPT-5.5), same tool layer (Composio), same context layer (Obsidian) — but the **brand on the harness commands a 2× price multiple at the customer level**. Hermes is positioned as the "self-evolving" premium tier; OpenClaw is the commodity tier. The repositioning has not appeared yet in either project's official marketing, but it is now public guidance from a sitting harness operator running real customers.

The [agent-harness-engineering essay](/posts/2026-05-09-agent-harness-engineering-the-discipline/) we published three days earlier mapped the technical convergence — every major coding agent ships the same twelve primitives. The Vasiles-Isenberg episode is the **commercial mirror**: the same primitives ship in every harness, but the *naming of the harness* now sets the price. Branding the substrate is the value-capture mechanism. This is the structure that played out in cloud (S3 vs Backblaze, same underlying object storage, different prices) and in databases (Snowflake vs Redshift, same query engine archetype, different multiples). It is now playing out in agent harnesses.

Hermes's premium position is propped up by three reinforcing assets that OpenClaw, despite its 371K-star advantage on GitHub, has not yet matched:

1. **Self-evolution.** [NousResearch ships an active-learning loop](https://hermes-agent.nousresearch.com) where the harness adopts new skills and rules over time without operator intervention. OpenClaw's skill system is human-curated.
2. **Mitosis cloning.** Nick's [April 17 tweet](https://x.com/nickvasiles/status/2048125488316248000) (284 likes) walked through OpenClaw and Hermes both supporting full replication of agent + computer + Obsidian vault. But Hermes's Orgo-coupled cloning runs sub-500ms; OpenClaw's takes longer in his reported runs.
3. **Gateway reliability.** The transcript is explicit: *"OpenClaw has a lot of gateway issues in my experience. Hermes is a lot better."* Gateway crashes are the operator's single most expensive failure mode because they break the customer-facing chat surface (Telegram, WhatsApp) without warning. Hermes's lower crash rate at the gateway layer is what carries the 2× pricing.

![Same agent body, different brand: Hermes at $10K/mo vs OpenClaw at $5K/mo, with self-evolution, cloning, and gateway reliability as the gap.](/post-images/2026-05-12-greg-isenberg-managed-agent-business-playbook/harness-pricing-tiers.jpg)

## The "use agents to build agents" loop

![One human with a Telegram phone connects to a meta-agent (Orgo Claw) which manages 27 customer agent VMs, all controlled from one chat.](/post-images/2026-05-12-greg-isenberg-managed-agent-business-playbook/orchestration-pattern.jpg)

The technical core of the episode is the demo where Nick installs a Hermes agent inside an Orgo cloud computer **using a different Hermes agent** that he orchestrates from a Telegram chat. Greg watches it happen live. The transcript captures the moment Nick types the install command from his phone:

> *"My agent is literally using an Orgo MCP to connect to my customers' agents that live on Orgo. And so what ends up happening is Orgo is like this workspace where my agent and other agents and myself can all collaborate on these computers where these agents live and get them set up and configured that way."*

He then queries the orchestrating agent:

> *"How many Orgo VMs do I have in my workspaces?"*

The agent replies with **27 VMs across his workspaces, all 27 running**, and a per-customer breakdown. This is the operational reality the playbook is built around: **one solopreneur orchestrating 27 customer-agent VMs via a meta-agent**, with Telegram as the control surface. At $5K MRR per customer (the floor), the implied ARR is $1.62M on 27 customers; at the Hermes-tier $10K MRR the same operator could clear $3.24M with the same human labor. Those numbers are not in the transcript — they are the obvious arithmetic Greg is selling. Nick neither confirms nor denies that he is at that revenue level; the only on-record claim is *"running this offer"* and *"working really well"*.

The pattern Nick frames as the philosophical core is **"more agents is the answer."** Concretely:

| Operator pain | Solution Nick demos |
|---|---|
| Hermes setup is hard | Use Claude Code or Codex inside the VM to install Hermes |
| Hermes config requires up-to-date docs | Give the builder agent Perplexity + Context7 + Exa + X MCP |
| Research tasks block the builder agent | Spawn five sub-agents — one each for Perplexity, Exa, Context7, Firecrawl, X MCP |
| Operator is away from the desk | Telegram-controlled meta-agent manages the whole fleet |
| Multi-tenant security | One Orgo workspace per customer; one VM per agent |
| Gateway crashes | A watchdog agent auto-restores |
| Customer-visible breakage | Mia (the customer's agent) emails the operator from her own inbox when a cron job breaks |

The watchdog-and-observability layer is the unsung work item. Without it, the unit economics collapse the first time a Telegram or WhatsApp gateway dies overnight and the customer wakes up to a dead assistant. The operator's competitive edge over a customer's in-house attempt is **specifically this perimeter**: the watchdogs, the email-from-agent alerts, the sub-second VM recovery. Not the agent itself, which is increasingly available as a one-click install.

## The competitive shape, post-publication

Public mid-market awareness of the agent-agency model has been thin until now. The operator community on X has been running it for months — Nick's own [May 5 post](https://x.com/nickvasiles/status/2052108413092528000) (22 likes, low visibility) gave the four-step recipe in compressed form:

> *"How to make money building an Agent Harness: — start by selling openclaw/hermes setups for SMBs — learn everything there is to know in your niche — build an 'ideal setup' of your agent and its computer environment — clone your setup anytime you get a new customer, removing setup time"*
>
> — [@nickvasiles](https://x.com/nickvasiles/status/2052108413092528000), May 5, 2026

The Isenberg podcast is the first version of that recipe to ship on a channel that reaches mainstream solopreneurs. Greg's audience is large (640K X followers, top-3 entrepreneurship podcast) and his content engine is precisely calibrated to **make obscure operator patterns into mass-market wedges**. He has done it before with the cohort of niche-newsletter operators (the *Boring Marketer* arc), with the Idea Browser pattern (turn subreddit complaints into product seeds), and with the productized-service movement broadly. This is the same play, applied to agent-agency.

The two-week window around the podcast shows Greg running the same thesis three times:

| Date | Format | Title | The thesis underneath |
|---|---|---|---|
| **May 9** | X post (32 likes, 9,149 views) | [*"Agent franchise kits"*](https://x.com/gregisenberg/status/2052988834928865000) | Find tiny AI agent business that works, package it, license to operators by niche |
| **May 9** | X post (40 likes, 8,564 views) | [*"A scorecard for rating AI agent businesses"*](https://x.com/gregisenberg/status/2052991022358593000) | Not all agent ideas are equal; rate by buildability × moat × wedge |
| **May 12** | Podcast (1,431 views in 21 hrs) | [*The $1M+ Solo AI Agent Business*](https://youtu.be/BI-MNjm1tTQ) | The end-to-end operator playbook, fully named stack, $5K-$10K MRR per customer |

This is a content-thesis sequence, not three separate ideas. The franchise-kit post seeds the concept. The scorecard post supplies the evaluation criteria. The Vasiles podcast supplies the **build instructions**. The implicit funnel is to send Greg's audience to **Idea Browser** (his startup-idea engine, [ideabrowser.com](https://www.ideabrowser.com/), where he is co-founder) to discover the niche, and to **Orgo** (Nick's company, where Nick is offering, per the YouTube description, to *"personally set up your Orgo in a 15-minute call"*) to ship the stack. Both businesses are real, both are referenced inside the episode, and the Greg-to-Nick relationship is openly disclosed.

🟢 **The result is that the underground operator pattern is now a public funnel.** Idea Browser captures the pre-build phase (niche selection). The podcast captures the build phase (stack + offer). Orgo captures the infrastructure phase (where the agents live). Composio captures the integration layer. AgentMail captures the inbox layer. The chain is end-to-end commercialized and the operator can plug in at any layer without learning the others. From a marketplace-design perspective, this is approaching the **Shopify-of-AI-agency** shape that no one has explicitly named yet but several teams are already racing for.

## What the playbook does not tell you

The episode is unusually honest by solopreneur-podcast standards, but four omissions are load-bearing for any operator considering the play.

🔴 **Acquisition cost is missing.** The transcript says *"get customers through content"* and *"content is the most leveraged thing you can do in 2026."* It does not say how many months of weekly content production it takes to land the first $5K MRR customer, what conversion rate to expect from Calendly bookings, or what the customer-acquisition-cost-to-LTV ratio actually looks like across Nick's claimed 27 customers. Late Checkout (Greg's agency) and Boring Marketer (his content product) both run on long-content-build motion that took years to mature. A new operator starting today does not have either advantage.

🔴 **Churn is not characterized.** The implicit promise of $5K MRR is annualized at $60K per customer per year. The lifetime value depends entirely on churn, and the episode is silent on it. Mid-market services contracts run somewhere between 12% and 28% annual churn depending on segment ([HubSpot benchmark, 2024](https://blog.hubspot.com/sales/customer-churn-rate)) — meaning $60K of annualized revenue is closer to $45K of expected LTV in the median case, before tax and OpEx. The "27 customers at $5K MRR = $1.6M ARR" math is sticker; the real number after churn is lower.

🔴 **The fulfillment ceiling is real.** Nick demos one operator managing 27 VMs. The episode does not say what happens at 50 customers, or at 100. Loom updates *"at 2 AM"* and *"under-48-hour turnaround"* are feasible at 27; at 100 they are not feasible for one human even with substantial agent automation. The model implicitly tops out before the operator becomes a $5M/year solopreneur — the realistic ceiling is somewhere in the $1.5M–$2.5M ARR band, which is still life-changing but is not the headline framing.

🔴 **The information-asymmetry premium just got compressed.** Underground operators selling at $5K MRR were doing so against customers who did not know what Hermes Agent was, did not know Composio existed, and did not know the marginal cost of an Orgo VM. The Isenberg podcast eliminates that asymmetry for the *operator side* of the market. Within twelve months, **the customer side will also know** — at minimum a slice of mid-market buyers will start asking "wait, why am I paying $5K/month for what is basically an Orgo subscription plus Composio plus Hermes?" The historical analogue is the drop-shipping market between 2018 and 2021: tutorial content commoditized the playbook, the price floor moved from $5K product launches to $500–$1,500 setups within twenty-four months, and the operators who survived were the ones who built **vertical-specific moat** (DTC brands rather than agencies, custom apparel rather than aliexpress repackaging).

The 2026 equivalent moats are not yet built. The candidates are vertical-specific Obsidian vaults (knowledge moats), custom skills shipped only to your customers (skill moats), watchdog/observability infrastructure trustworthy enough that customers pay for *reliability*, not for *agents* (reliability moats), and brand authority in a specific niche (audience moats — what Greg is himself building, and what most operators will not be able to replicate at his scale). Operators who treat the Isenberg playbook as the destination rather than the starting line will be the ones displaced when the price floor moves.

## The discourse, 24 hours in

The first-day signal on X is small but the *shape* of the conversations is informative. The thread had 5 organic replies by 21 hours after launch; none were hostile, two were affirmative, one was a skeptic asking the right question. The skeptic reply is the one that matters:

> *"Managed AI agents for business sounds compelling until you start thinking about error handling, auth, and the edge cases that eat up 80% of the build time. Curious how you address that."*
>
> — [@Chahatxsharma](https://x.com/Chahatxsharma/status/2054262000000000000), May 12, 2026

This reply names the gap Nick spends nine minutes of the episode on (watchdogs, observability, scope-limiting, customer-facing email alerts) but did not signpost in the X thread. The thread's marketing surface (the 10 numbered bullets) emphasizes the upside and elides the reliability work. Inside the podcast itself, the reliability work is half the actual playbook. Operators who only read the thread will miss it; operators who actually watch the 47 minutes will not. That asymmetry is itself a fulfillment moat for whoever takes the harder content seriously.

In the broader agent-agency discourse the same week, three pieces of independent corroboration matter:

> *"Yesterday, I set up Hermes and Paperclip. I'm testing an AI Agent Agency team. My Hermes agent (sort of like OpenClaw, but less glitchy) manages a team of subagents via Paperclip."*
>
> — [@caelanhuntress](https://x.com/caelanhuntress/status/2052402935290318000), May 7, 2026

The Huntress tweet, posted five days before the Isenberg podcast, is the *"already running this offer in private"* signal — an operator independently arriving at the same Hermes-manages-subagents architecture without coordination. The Paperclip layer is a different sub-agent orchestrator from Nick's Orgo-MCP-via-Telegram setup but the shape is identical. Convergence across operators, not coordination, is the evidence the architecture is correct.

A second independent signal came from Tokyo. On the same day Greg shipped the podcast, Japanese tech personality [Takafumi Horie](https://x.com/takapon_jp) launched a productized service called **AI Agent Agency (AAA)** for staffing-shortage companies:

> *"🔔お知らせ🔔 人手不足に悩む企業向けにAI社員を構築提供、「AIエージェント構築代行サービス(AAA：AI Agent Agency)」開始のお知らせ"*
>
> — [@horiemonai](https://x.com/horiemonai/status/2054287000000000000), May 12, 2026

(Translation: *"Notice. For companies suffering from staffing shortages, announcing the launch of an AI-employee-build service called AAA (AI Agent Agency)."*) The Japanese launch is a separate, regulated, branded productization aimed at the macro pain (Japan's labor-shortage crisis), which is a much harder problem-fit than Nick's US-SMB pitch but the same product shape. Two simultaneous launches on different continents on the same day is not coordination; it is the **shape of a real market arriving on schedule**.

The skeptic position is also represented. The harness-engineering Twitter critique came from a Chinese-language thread the same week:

> *"harness工程的价值是产生规范数据给llm训练用。而真llm的核心是'能动性'而非skill/workflow，老用workflow思想要不得，灵活性极差。"*
>
> — [@xsser_w](https://x.com/xsser_w/status/2053491000000000000), May 10, 2026

(Translation: *"The value of harness engineering is producing structured data for LLM training. The core of a real LLM is 'agency', not skill/workflow. Repeatedly using workflow-thinking is unacceptable; it is extremely inflexible."*) The argument is that the entire Isenberg-Vasiles playbook is workflow-coded — that it bolts skills onto agents in a deterministic shape that future model generations will dissolve, leaving the operator's moat structurally short-dated. The position is contestable (operators counter that human-supervised workflows are the *only* reliable shape for production deployment in 2026) but it is the cleanest articulation of the bear case in the public discourse this week.

![Five independent operator signals on the agent-agency thesis in the May 7-12 window: convergence, not coordination, across three continents.](/post-images/2026-05-12-greg-isenberg-managed-agent-business-playbook/discourse-arc.jpg)

## The Orgo side of the trade

![Where the customer's $5,000 lands: thin slices to Orgo, Composio, SaaS tools, OpenAI tokens, with ~$4,380 to operator margin.](/post-images/2026-05-12-greg-isenberg-managed-agent-business-playbook/revenue-split.jpg)

Nick is the founder of [Orgo](https://orgo.ai/), the cloud-computer-for-agents company. The product positioning on the home page is exact: *"Orgo is computers for AI agents. The developer platform for computer-use agents with sub-500ms boot time."* The episode is, among other things, a 47-minute Orgo demo. That is not a critique — it is openly disclosed, and the episode would be less useful if Nick pretended otherwise. But it does set up a question the playbook does not answer: **what fraction of the captured value lands with Orgo vs with the operator?**

The implicit revenue split, if you accept the episode's numbers, is roughly:

| Layer | Recipient | Approx. share of $5K/month |
|---|---|---|
| Orgo VM (1-3 VMs per customer, ~$30/VM enterprise) | Orgo | ~$50-90 |
| Composio (~$100/month per active customer at scale) | Composio | ~$100 |
| AgentMail / Granola / Trello / Loom / Superhuman | Several SaaS vendors | ~$100-200 |
| Model tokens (GPT-5.5 default, paid plan) | OpenAI | ~$200-400 |
| **Operator margin (gross)** | Solopreneur | **~$4,200-4,500** |

The vast majority of customer dollar value flows to the operator. This is good for the operator, neutral for Orgo (who collects per-VM revenue at scale), and structurally bad for the harness projects themselves (NousResearch's Hermes is MIT-licensed and free; OpenClaw is MIT-licensed and free). The harnesses capture no direct economics; they capture brand value that the operator can resell. **This is exactly how Linux distributions worked in the 2000s**: Red Hat captured none of the kernel's value, but the value of Red Hat the distribution was the *enterprise wrapper* — support, certification, packaging, training. The agent-agency operator is the Red Hat of the harness. The harness itself is the kernel.

The interesting structural question is whether Hermes or OpenClaw eventually move to capture some of this value directly — by offering a managed-Hermes-for-SMB hosted service, by introducing per-agent pricing for commercial use, or by partnering with infrastructure providers like Orgo on a revenue-share. NousResearch shipping a hosted-Hermes Cloud product would compress operator margin overnight. The fact that it has not happened by May 12, 2026, is itself a signal: the harnesses are betting that the *brand* of being free-and-open will continue to outweigh the captured-revenue alternative. If that bet flips, the agent-agency operator's economics flip with it.

## What it would take for this to be the next "drop-shipping"

The hostile read on the Isenberg-Vasiles episode is that this is the 2026 version of the e-commerce drop-shipping content economy. A recipe shipped at scale, attractive to operators who under-estimate execution difficulty, and ultimately good for the picks-and-shovels businesses (Orgo, Composio, Idea Browser) but unevenly good for the operators who attempt it. The friendly read is that this is the 2026 version of the *productized service movement*. A legitimate operating model that built real businesses like [Designjoy](https://www.designjoy.co/), [Pineapple](https://www.pineapple.dev/), and dozens of niche SaaS-replacement agencies that quietly cross $1M ARR.

The diagnostic between the two reads turns on three measurable things:

| Diagnostic | Drop-shipping outcome | Productized-service outcome |
|---|---|---|
| Customer LTV after 12 months | under $15K (high churn, no moat) | over $45K (real workflow embedded) |
| Operator's content-to-customer conversion | under 1 in 5,000 viewers | over 1 in 1,000 viewers (paid funnel works) |
| Margin pressure at month 18 | -40% (commoditized) | -10% (vertical moats hold) |
| Price floor at month 24 | $500-1,500/month | $4,000-7,500/month |
| Survivor rate at month 24 | under 10% of operators | 30-50% of operators |

The data points to watch over the next quarter are: (1) whether the operator-side X discourse stays high-quality or fragments into "I made $10K my first week" posts that signal NFT-style speculation; (2) whether Hermes or OpenClaw introduce commercial hosting (which would compress the operator margin); (3) whether Composio's auth layer becomes the canonical bottleneck (which would let Composio extract pricing power); (4) whether the customer side starts to publicly compare operator pricing (which would commoditize the offer); (5) whether the first real customer-side success stories — *"my law firm cut associate hours by 40% with this agent"* — appear with verifiable specifics rather than testimonial-only marketing.

The 90-day signal is the X discourse itself. If by mid-August 2026 the agent-agency content fills with operator pricing comparisons, customer-side complaints about identical tooling at differentiated prices, and harness-project hosted commercial offerings, the commoditization arc is real and the price floor moves down. If instead the discourse fills with vertical-specific success cases (the canonical shape: *"the medical-records-summarization Hermes agent for orthopedic clinics in Texas, $7,500 MRR, 14 customers, 2% monthly churn"*), the productized-service arc is real and the operators win.

🟢 **The most useful read of the Isenberg episode is that it tells you which agents are actually being sold and at what price.** That information is in itself worth the 47 minutes, independent of whether the playbook will work for any specific operator who tries it. The market just got mapped in public. Whether mapping the market also commoditizes it, or whether mapping it grows it, is the bet the next six months of operator behavior will settle.

The cleanest version of the bet is the one Nick himself wrote five days before the podcast, in the obscure low-engagement tweet that previewed the entire episode in fifty-eight words:

> *"How to make money building an Agent Harness: start by selling openclaw/hermes setups for SMBs. Learn everything there is to know in your niche. Build an 'ideal setup' of your agent and its computer environment. Clone your setup anytime you get a new customer, removing setup time."*
>
> — [@nickvasiles](https://x.com/nickvasiles/status/2052108413092528000), May 5, 2026

The episode is the long-form. The tweet is the thesis. Whether to take the trade depends on what you believe about how fast the next twelve months commoditize the cloning step.

## Sources

- [Greg Isenberg — The $1M+ Solo AI Agent Business (Full Course) [YouTube, 47 min, May 12, 2026]](https://youtu.be/BI-MNjm1tTQ)
- [Greg Isenberg — launch thread (X, May 12, 2026)](https://x.com/gregisenberg/status/2054261832718889216)
- [Nick Vasiles — How to make money building an Agent Harness (X, May 5, 2026)](https://x.com/nickvasiles/status/2052108413092528000)
- [Nick Vasiles — agents can clone themselves (X, April 17, 2026)](https://x.com/nickvasiles/status/2048125488316248000)
- [Nick Vasiles — long-horizon agent tasks (X, April 29, 2026)](https://x.com/nickvasiles/status/2050988283415453000)
- [Orgo — Computers for AI Agents (product site)](https://orgo.ai/)
- [Orgo on X — enterprise-ready infrastructure (Apr 21, 2026)](https://x.com/orgo/status/2049615870310301472)
- [Greg Isenberg — Agent franchise kits (X, May 9, 2026)](https://x.com/gregisenberg/status/2052988834928865000)
- [Greg Isenberg — outcome bounties / scorecard (X, May 9, 2026)](https://x.com/gregisenberg/status/2052991022358593000)
- [NousResearch/hermes-agent (GitHub, 146,583 stars)](https://github.com/NousResearch/hermes-agent)
- [openclaw/openclaw (GitHub, 371,224 stars)](https://github.com/openclaw/openclaw)
- [nickvasilescu/hermes-desktop-os1 (GitHub, May 9, 2026)](https://github.com/nickvasilescu/hermes-desktop-os1)
- [ComposioHQ/composio — 1,000+ toolkits (GitHub)](https://github.com/ComposioHQ/composio)
- [Composio docs](https://docs.composio.dev)
- [AgentMail — emails for agents](https://agentmail.to/)
- [Obsidian — Markdown knowledge base](https://obsidian.md/)
- [Anthropic — Claude Code](https://www.anthropic.com/claude-code)
- [OpenAI — Codex CLI / Codex desktop](https://openai.com/codex/)
- [Caelan Huntress — testing AI Agent Agency team via Hermes + Paperclip (X, May 7, 2026)](https://x.com/caelanhuntress/status/2052402935290318000)
- [Horie Takafumi — Japanese launch of AI Agent Agency (AAA) service (X, May 12, 2026)](https://x.com/horiemonai/status/2054287000000000000)
- [@xsser_w — harness engineering ≠ agency, skeptic view (X, May 10, 2026)](https://x.com/xsser_w/status/2053491000000000000)
- [The Deep Feed — Agent Harness Engineering, the discipline (May 9, 2026)](https://www.thedeepfeed.ai/posts/2026-05-09-agent-harness-engineering-the-discipline/)
- [Idea Browser — Greg Isenberg's startup-idea engine](https://www.ideabrowser.com/)
- [Granola — meeting note tool used in the stack](https://www.granola.ai/)
- [Trello — project management tool referenced as customer-facing kanban](https://trello.com/)
- [Z.AI / GLM 5.1 — open-source model recommended for lighter tasks](https://z.ai/)
- [Cline — `use_subagents` parallel spawning](https://github.com/cline/cline)

---

Canonical: https://www.thedeepfeed.ai/posts/2026-05-12-greg-isenberg-managed-agent-business-playbook/
Site: https://www.thedeepfeed.ai
Full corpus: https://www.thedeepfeed.ai/llms-full.txt