How to Run a One-Person Company With a Team of AI Agents

AI agents can help a solo founder research, prepare drafts, and maintain recurring internal work. That does not establish that a business can run unattended or earn revenue without people. The useful question is narrower: which repeatable job can an agent complete well enough to reduce your workload?

What an agent-run company actually means

The founder still owns priorities, customer commitments, spending, and final approval. An agent receives inputs, uses permitted tools, produces an artifact, and returns evidence for review. Scheduling starts the work; it does not make the output correct. A second agent can check a draft, but its agreement is not independent proof.

Use role files as operating instructions

A folder and a Markdown file are a simple way to package a role. Record its objective, permitted sources, tools, output format, acceptance criteria, and actions requiring approval. Keep credentials outside the file. A written instruction is not an access control: enforce restrictions through the tool or account permissions too.

Six roles you can add as needed

• Research: collect sources and identify unanswered questions.
• Content: draft from approved facts and a defined brief.
• Leads: classify inbound inquiries and draft replies for review.
• Operations: prepare checklists and flag overdue work.
• Finance and documents: prepare invoice drafts and reconcile records for human checking.
• Review: check claims, completeness, and source references.

These are suggested responsibilities, not a required six-agent organization. Start with the role your workload actually needs.

Schedule outputs, not activity

Use a scheduler such as cron or a supported application automation. Choose the time zone, required inputs, output location, maximum duration, and spending limit. An example schedule is a morning briefing, a midday check of new inquiries, and a weekly newsletter draft. A periodic health check should notify you of failures or required decisions, rather than sending empty updates. Sleeping computers and expired credentials can stop local jobs; test those conditions before relying on the schedule.

Build one workflow before adding a team

1. Pick one frequent task and save examples of acceptable human work.
2. Give the agent a bounded input set and a draft-only output destination.
3. Run it manually and check every factual claim against its source.
4. Add a separate review pass and retain the founder’s final approval.
5. Schedule only after repeated successful runs; inspect missed runs, duplicate outputs, and costs.

MCP can connect an AI application to tools and data. It is a connection protocol, not a guarantee that a connector grants the right permissions or handles your data appropriately. MCP

A newsletter task you can test

Every Friday at 09:00 Europe/London, use the approved source folder to draft a 500-word newsletter. Include a source link for each factual claim and a list of unresolved questions. Save the draft to the review folder. Do not send email or publish. If a source is unavailable, report the gap instead of inventing a replacement.

This is a proposed task specification, not a claim that a particular scheduler accepts this prose. Configure the equivalent schedule and permissions in your chosen tool.

Handle failures and permissions explicitly

Give each run a unique identifier and log inputs, output, duration, tool calls, and cost. Prevent overlapping runs and duplicate sends. Retry only operations that are safe to repeat, and cap attempts. Require explicit approval for customer messages, payments, publication, and destructive changes. A reviewer should see the artifact and sources before approving, not just a success notification.

Try the workflow in Eigent

Eigent is an open-source desktop workspace for agents and connected tools. You can organize context in a Space and inspect work through Sessions and Tasks. Its website describes automations, source inspection, and approval controls; verify the behavior of the version and setup you use. Local execution does not keep every request on-device: cloud models and external connectors can transmit information to their providers. Choose local models and compatible tools when the task requires that boundary. Eigent · GitHub

Decide whether to expand

For several runs, record accepted outputs, factual corrections, review minutes, failed or duplicate runs, and total model/tool cost. Compare those results with doing the job manually. Add another role only when the first workflow reliably reduces total effort. Read the AI office agents comparison for other approaches, then download Eigent and start with one reviewable draft. AI agents · Eigent