Software-enabled services
Sell the finished outcome. Agents do the delivery work behind a human relationship.
Agent-Run Business / Capability Audit
The influencer thesis is more right than its critics allow: software is cheap now, agents are becoming the users, and a single operator can run what used to need a team. The thesis is also incomplete in one specific way, and the gap is measurable.
The argument on offer
The clearest version of the case comes from the Build Great Products channel, in a video published October 21, 2026 under the title "Stop building SaaS. Build one of these 8 instead."
The reasoning runs in three steps. Building software has become cheap, because tools like Claude Code and Codex make cloning a generic product trivial. Agents are becoming users of software, not just builders of it. And pricing is shifting from seats toward outcomes. Put together, those three moves strip the defensibility out of generic SaaS.
What remains defensible is your expertise, your taste, your data, and your audience. Not your features.
That is a sound read of the market. It is also where most of this genre of advice stops: with a list of shapes to build, and no account of what it costs to keep one running once agents are doing the work unattended.
The eight shapes
These are the eight categories named in the video. The descriptions are the plain reading of each, since the source lists the titles without elaborating on them individually.
Sell the finished outcome. Agents do the delivery work behind a human relationship.
One industry's rules, formats and exceptions wrapped around an agent loop.
CLIs, APIs and MCP servers. No interface, consumed by other agents.
A plugin that lives inside somebody else's agent, plus a dashboard you own.
Judgment and craft as the moat, in a market too small for a funded competitor.
Always running, acting in the background on a person's behalf.
A dataset nobody else holds, sold through whatever surface fits.
Give away the engine, charge for the thing around it.
The source material
Each one sells a different version of the agent-run company: a shape to build, an income figure to hit, and a ranking to obey. Worth watching in that order.
The structural argument. Software got cheap, agents became the users, pricing moved to outcomes, so generic SaaS stopped being defensible. Supplies the eight shapes above.
The income framing, and the one to hold against the production numbers in the next section. A monthly target plus zero headcount is a claim about reliability as much as revenue.
The tier-list framing. Independent rankings of the same question tend to land in one place: AI SaaS near the bottom because building the app is now trivial while distribution stays brutal, with consulting, audits and lead generation rated most durable.
A note on method. Only the first of the three could be retrieved directly; the other two return a client error from YouTube and may be restricted. They are credited here as source material, and their framings are described from their titles and channels rather than from their contents. Nothing in the eight shapes or the vendor table below rests on them.
The part that gets left out
The platforms can do this. That is not in question anymore. What the production data says is that the failures cluster somewhere else entirely, and they cluster in a way that matters more to a one-person operation than to a company with staff to catch mistakes.
Share of agentic AI projects Gartner forecasts will be canceled by the end of 2027, attributed to cost, unclear value, and inadequate risk controls. A forecast, not a measured rate.
Multi-agent failures that Microsoft Research traced to agent boundaries, the handoff points between agents, rather than to any single agent's reasoning.
Increase in attack surface per autonomous agent relative to a human user, in Cisco's analysis, because agents inherit permissions and credentials in trust-by-default mode.
Vendors Gartner estimates are delivering genuine agentic capability, out of the thousands claiming it. The rest is rebranded chatbots and RPA.
Read those four numbers together and a pattern falls out. The model is not what breaks. The seams break: handoffs between agents, permissions nobody scoped, costs nobody capped, and a vendor you picked who turns out to be one of the thousands rather than one of the hundred.
A solo operator has no second pair of eyes. The control layer is not overhead for them. It is the only thing standing between an unattended agent and an unrecoverable afternoon.
The practitioners who make this work converge on the same unglamorous habits. Scope autonomy narrowly and widen it only against proof. Design the human verification gate before launch rather than after an incident. Cap what a single task may spend before it escalates, with $500 per task offered as one recommended production ceiling. Name one accountable owner per agent with rollback authority, which for a solo operator means admitting that the owner is always you.
The mapping
| The vision | What executing it actually requires | Vendors capable today | Verdict |
|---|---|---|---|
| Software-enabled servicesSell the outcome, agents deliver | An internal agent fleet plus a human client relationship. Reliability matters only to you, not to a buyer inspecting the stack. |
|
Buildable nowThe most proven shape on this list. You absorb the failures before a customer sees them. |
| Vertical agent harnessesOne industry, mastered end to end | A real harness: agent loop, context compaction, tool approval, subagents, session state. Plus the vertical's rules carried inside the product, and a retrievable log for every action. |
|
High skill barCapability is there. Harness engineering is ongoing work, not a one-time build. |
| Headless AI softwareCLIs, APIs, MCP servers | Publish tools other agents consume. No interface to design, no onboarding to write, and distribution through an existing protocol. |
|
Buildable nowLowest capital cost of the eight. MCP sits under every major harness with 200+ servers active. |
| Agent-native systemsPlugin inside someone else's agent | Distribution through a platform you do not control, and a dashboard you do. Your roadmap becomes a function of their roadmap. |
|
Platform riskOpenAI launched AgentKit in Oct 2025 and deprecated it on June 3, 2026, shutting down Nov 30. Build on a host, inherit its reversals. |
| Taste-led niche productsJudgment as the moat | Almost no agent infrastructure. The scarce input is your own discrimination, which no vendor supplies. |
|
Buildable nowThe least vendor-dependent shape here, and therefore the least exposed to anyone's deprecation notice. |
| Ambient consumer agentsAlways on, acting for a person | Persistent scheduled execution, durable memory, device and context access, voice, and a consumer's trust with their credentials. Every failure is visible to a stranger. |
|
Weakest linkFull unsupervised autonomy is characterized as mostly demos. Consumer-facing is where trust-by-default permissions hurt most. |
| Proprietary data productsA dataset nobody else holds | Acquisition, cleaning and a query surface. Agent autonomy is nearly irrelevant to whether this works. |
|
Buildable nowThe moat is the data, not the agent. Which also means no vendor can take it from you. |
| Open source core productsFree engine, paid surround | A community and a credible monetization edge. Neither is a technical problem, and neither is something a platform sells you. |
|
Assembly requiredTechnically the easiest, commercially the hardest. Check each license before you build a business on it. |
The rows are the eight shapes from the Build Great Products video, which is the only one of the three sources whose contents could be read. If the Ramonov ten or the Herk ranking name a business model that none of these eight covers, this table needs rows, and the verdict column is where to look first.
What follows from it
One person can run a business with agents. That is now true, and the table above shows five of the eight shapes are buildable today with products you can start on this week.
What is not true is the implied promise underneath the genre, which is that the agents handle it and you collect. The credible deployments sit at supervised autonomy, where the agent acts and flags exceptions and a human monitors. Full autonomy with nobody watching remains mostly demos. Across thousands of deployed agents, task success has been measured at 56.6%, with a 37-point gap between benchmark performance and reality.
None of that is an argument against building. It is an argument about where the work actually is. The influencers are right that the moat moved off the code. They are wrong about where it landed.
It landed on the control layer. Scoped permissions, spend caps, verification gates, retrievable logs, and one named owner. That is the real differentiator between the operators who ship and the 40% who cancel.
Which is a strange and useful conclusion, because the control layer is the one part of this that no vendor on the table above will hand you finished. It is the part you have to own.
The domain under this page is for sale
An agentpreneur is what the three videos above are describing without quite having a word for it: one operator, no employees, a business run by agents. The term is already in use in newsletters and communities. The adjective form is not yet spoken for.
This page is the proof of concept. It is what the domain looks like pointed at the category, on day one, with no build required.
Sources
Two of the three videos submitted for this review could not be retrieved. Figures described as forecasts are forecasts and not measured outcomes. Vendor verdicts are an assessment, not a vendor claim, and the agent tooling market has been revising itself roughly quarterly.