Case study · Our own company
How a one-person company runs on AI agents
The founder organizes the company around 14 specialized agent roles in four areas. They do focused work, report to the founder directly, and wait for approval before spending money or acting outside their role.
Four areas. One person accountable.
There is no boss agent between the founder and the specialized agents. Each role has its own instructions and access; the founder receives the reports and makes the decisions.
Founder
Direct supervision · no boss agent
Growth
- Offer & acquisition agents for training, consulting, and software
- Monthly offer-experiment agent: up to two offers and three inexpensive acquisition tests
Software products
- One agent per SaaS product
- Two niche-radar agents on different AI engines
Operations
- Weekly review agent (read-only)
- Monthly finance review agent (never moves money)
Systems & quality
- IT agent
- Code-quality agent
- Automation reliability agent
Shared tools make the work visible
Slack: the conversation
Each agent has a channel for short reports, questions, and approval buttons. The founder can trigger a mission with a message or respond to a scheduled report.
The ERP: the operating record
The company’s own ERP holds clients, quotes, and invoices, plus the status of each agent run: succeeded, failed, blocked, or late.
The real work tools
Agents work with scoped code repositories, folders, calendar, and email. They use the same operational sources the company relies on.
Written instructions
Each role has a mission, sources, permitted actions, and escalation rules. Access follows the role; it is not shared indiscriminately across agents.
The mission loop
Every mission has a trigger, a scope, and a recorded outcome.
- 1
Trigger
A schedule or a message in the agent’s Slack channel starts a mission.
- 2
Context
Memory, written instructions, and sources within the agent’s scope.
- 3
Work
Do the task or delegate to execution agents with their own access.
- 4
Report
A few lines in Slack and a run status recorded in the ERP.
- 5
Decision
Spending and anything outside the role wait for the founder. Other work is logged.
Autonomy has explicit boundaries
Done alone, then reported
Tested internal fixes can be merged when CI is fully green. Agents may restart a failed automation or clean documentation and temporary files within their scope.
Approved in Slack
Any spending; any new message an agent writes to a client, prospect, or partner; and anything outside the agent’s scope, including databases, secrets, or another app’s code. Routine messages from approved templates, such as booking reminders and product emails to a SaaS product’s own users, go out on their own. Client projects never ship without human acceptance testing.
Limits that stay in place
Review agents prepare; they do not decide. Agents run on schedules within AI service usage limits, rather than as a round-the-clock team. An independent, non-AI check verifies morning and evening that each automation ran.
AI agents in production: who does what
Measured on October 7, 2026: 14 orchestrator agents configured (13 on Claude, 1 on Codex), 21 scheduled jobs, and 300 runs between October 2 and October 7, of which 298 succeeded (99.3%), with 15,921 tool calls.
| Agent | Mission | Does on its own | Submits for approval |
|---|---|---|---|
| Marketing (training and consulting) | Traffic and bookings for two websites | SEO and content updates; sends emails whose text is in an approved plan, capped weekly | Any new message text and every reply |
| Marketing (software and AI agents) | Traffic and bookings for four websites, including this one | SEO and content updates | Messages to third parties |
| Prospective | Up to two new service ideas and three acquisition ideas a month | Researches and measures | Each idea before it is tested |
| SaaS product agents | Growth and maintenance of one product each | Deploys on green tests, updates its database, sends product emails to its users | Spending; on one product, features requested by users |
| SaaS radar (Claude and Codex) | Find niches, open a landing page, then build the product | Deploys, creates free-tier resources and prices | Buying a domain, moving to a paid plan, refunds |
| Business review | Weekly review of the ERP | Reads and cross-checks sources | Everything: it writes nothing |
| Finance | Monthly review of both companies and the household | Analyzes and recommends | Any payment, transfer, filing or email |
| IT hygiene | Simpler, safer systems | Docs, temporary files, merged branches | Production, databases, secrets, other agents |
| Code review | Reviews one project each run | Merges internal fixes on green CI | Client projects: never merges |
| Repair | Fixes failed automations after the watchdog | Restarts a service, reinstalls a timer, fixes an automation script | App code, databases, secrets, spending |
How we run these agents on Claude → · AI agent development for your business →
What those roles can work on
Illustrative examples drawn from each agent’s instructions, not execution logs.
A growth experiment proposal
Outline a niche offer, the audience, a small acquisition test, and the signal worth checking. Present it for a decision before spending or publishing commitments.
A product improvement
Investigate a scoped issue, propose a fix with tests, and check CI. An internal change may be merged only within the role’s written permissions.
A weekly operating review
Read the ERP, email, and calendar, identify items needing attention, and prepare a prioritized list with sources. Leave the decisions to the founder.
A failed automation
Use the independent check’s failure report to diagnose the cause. Apply a permitted reversible repair or ask for approval when the problem crosses the role’s boundaries.
What we can set up for you
Start with one or two well-chosen roles connected to your tools, with your approval rules. The right starting point is your bottleneck—not a copy of this organization.