How to Run a One-Person Company With AI Agents

The pitch is everywhere: one founder, a team of AI agents, a company that runs while you sleep. The serious guides say something more modest. Eigent's guide to running a one-person company with agents states plainly that agent help "does not establish that a business can run unattended or earn revenue without people." IV Consulting's playbook opens with "More agents is not the answer." Even Taskade's long survey of solo businesses, written to sell an agent workspace, concedes that most solo founders earn modest incomes and that the million-dollar examples are outliers.

All three sell tools or services, so read them as practitioners' advice rather than evidence. Their overlapping recommendations provide a starting checklist. Adapt it to the task, then measure whether it reduces the operator's work.

The human keeps four jobs

Agents now draft, research, sort, reconcile and write code well enough to take real work off one person. What they cannot take is the founder's judgment. Taskade frames the founder as the board of an agent organization, keeping strategy, taste and customer relationships. Eigent adds priorities, customer commitments, spending and final approval. Put together, the human keeps four jobs:

1. Deciding what to work on and for whom.
2. Owning every promise to a customer, including what gets sent and when.
3. Approving money and anything that cannot be undone.
4. Judging quality, because when everyone can produce, as Taskade puts it, "production stops being the advantage."

Everything else is a candidate for an agent, one job at a time.

Build in layers, in order

IV Consulting describes six layers and insists they are "a sequence, not a menu":

Layer	What it is	Where it breaks
Foundation	Clean data, written procedures, one source of truth	Agents act confidently on messy data
Single-job agents	One narrow agent per repetitive task	Scope creep
Orchestration	A thin router that hands work between agents	Context lost at the handoff
Memory	Shared facts the agents read and write	Old prices or policies treated as current
Monitoring	Alerts on business results	Failures that look like success
Human checkpoints	Approval before irreversible actions	Approving on autopilot

Check the source data before connecting the agents. An agent using a stale price list can produce plausible quotes with the wrong prices.

Write each agent a role file

Eigent's most practical idea is to give every agent a short written file: its objective, the sources it may use, its tools, its output format, what counts as acceptable, and which actions need approval. Typical first roles are research, content drafting, sorting inbound leads, operations checklists, invoice drafts and a separate reviewer that checks claims against sources.

Then comes the line most guides leave out: "A written instruction is not an access control." If an agent must never send email, take away its permission to send email. Enforce that restriction in the tools or service. Credentials stay out of the role file, each agent gets its own limited account, and anything the agent could break is backed up before it starts.

Start with one workflow, in shadow mode

Both IV Consulting and Eigent recommend the same start:

1. Pick one frequent, low-risk task and save examples of good human work.
2. Give the agent a bounded set of inputs and a draft-only output folder.
3. Run it by hand, and check every factual claim against its source.
4. Run it in shadow mode for a week: it proposes, you approve everything.
5. Schedule it only after repeated clean runs, with a time limit and a spending limit for each run.

IV Consulting's plan takes 90 days: two weeks on data and procedures, a month on the first agent, a month on the second agent and its handoff, and the last weeks on monitoring. Add a second agent only when the first has clearly reduced your total effort. Eigent measures that directly: accepted outputs, corrections, review time, failed or duplicate runs, and model and tool costs, compared with doing the task yourself.

The two-question checkpoint test

Where should a human approve? IV Consulting reduces it to two questions. Can I undo this in one click? Does a mistake reach a customer, or does it stay internal?

• Reversible and internal: consider automatic processing where a mistake is contained. Ticket tags and draft replies may qualify; confidential or high-stakes work still needs its own review.
• Irreversible or customer-facing: a person approves. Sending the email, paying the invoice, publishing, deleting.

Keep the checkpoints few, or they become a reflex. IV Consulting's warning is the agent that is right ninety-nine times, so the hundredth gets approved without reading. A good checkpoint shows the evidence, not just the output: the source behind a claim, the diff behind a change, and what else depends on the thing you are about to delete. For example, a cleanup procedure should check whether an old release directory is still referenced by a live service before proposing deletion. Age alone is not enough evidence.

Monitor the result of the work

IV Consulting names three failures that a smarter model will not fix, because they are about structure, not intelligence:

• Handoffs, where part of the context does not travel from one agent to the next.
• Memory drift, where an agent keeps using a fact that has changed.
• Silent errors, where the log says "success" while the result gets worse. "A silent failure looks exactly like success until a customer complains."

So monitor what the business cares about: replies actually sent, invoices actually paid, pages actually loading, not whether a job exited cleanly. Eigent's operating rules help here too. Give every run an ID and log its inputs, outputs, tool calls and cost. Prevent overlapping runs and duplicate sends. Retry only actions that are safe to repeat, with a cap. Then spend fifteen minutes a week reading the logs and the review queue.

What it costs, and what it earns

Software subscriptions are only one part of the cost. Taskade estimates a typical solo stack of subscriptions at roughly $85 to $200 a month, and an agent setup at a few hundred dollars a month. These are the vendor's own estimates. The real cost is the one IV Consulting names: oversight, testing and the occasional rebuild when a layer breaks. Compare model spending with accepted results, including the cost of retries and review.

The income side needs the same honesty. The solo founders everyone cites, such as Pieter Levels or Danny Postma, report their own revenue, and Taskade notes that the median solo founder earns far less. Agents lower the cost of producing work. They do not create demand. Use the time measurements to decide whether to continue. If an agent saves drafting time but creates an equally large review queue, change the workflow before scheduling more runs.

If you want help designing the first workflow, its permissions and its checkpoints, get in touch.