Basics
| Item | Details |
|---|---|
| Founder | Ben Broca (@bencera) |
| Product | Polsia — an AI Agent-operated platform |
| Team size | 1 person, zero employees |
| Revenue | $500K/month MRR, reached in 3 months |
| AI stack | Claude Opus 4.6 (“AI CEO”) + Claude Code + OpenAI Agents SDK + MCP protocol |
| Tech stack | Node.js, BullMQ, Redis, Render, Neon (Postgres), Stripe |
| Source | Indie Hackers original post / Fortune |
Storyline
Background: Ben Broca is an indie hacker who spotted a counterintuitive opportunity in the AI Agent wave — instead of using AI to assist human work, let AI Agents directly operate the entire business. He decided to build a platform with Claude Opus 4.6 as the “AI CEO,” autonomously managing everything from acquisition to delivery.
Turning point: The pivotal moment was the maturation of the MCP (Model Context Protocol). MCP lets AI Agents directly invoke external tools — Stripe payments, database queries, email — without humans in the middle. Ben designated Claude Opus 4.6 as the top-level orchestrator, managing a set of specialist Agents (marketing Agent, support Agent, finance Agent) through the OpenAI Agents SDK — an “AI managing AI” hierarchy. The architecture lets a single human handle thousands of concurrent customer requests.
Outcome: $500K/month MRR within 3 months. But it also exposed a serious risk — 20 Stripe disputes silently failed because no human was supervising, and it wasn’t caught until Ben manually checked. The case is a perfect illustration of the ceiling on AI autonomous operations: execution can be fully automated, but exception handling must have a human backstop.
Workflow Breakdown
Minimum loop analysis
| Step | What he did | AI involvement | Notes |
|---|---|---|---|
| 1. Think it through | Ben defined the “fully autonomous operations” product vision and the scope of Agent replacement | Medium | Direction is human; Claude assists in clarifying boundaries |
| 2. Write a spec | Used Claude to generate a behavior spec for each Agent (role / permissions / triggers / exception handling) | High | Every Agent has an explicit spec.md |
| 3. Implement | Claude Code generates Node.js code from the spec; BullMQ for task queues | High (~90%) | Core code is almost entirely AI-generated |
| 4. Validate | Validated with real Stripe transaction data, monitored Agent behavior logs | Medium | Validation is the weakest link — 20 silent-failure disputes are the cost of insufficient validation |
| 5. Iterate | Adjusted Agent behavior parameters based on logs and customer feedback | High | Claude analyzes logs and auto-suggests improvements |
Tool stack
- Claude Opus 4.6 (“AI CEO”): top-level orchestrator, responsible for task allocation, prioritization, and exception escalation. A canonical L2 (Agent-of-Agents) multi-Agent orchestration setup.
- OpenAI Agents SDK: the Agent execution framework — defines each Agent’s tool permissions and behavioral limits.
- MCP protocol: lets Agents directly operate Stripe, the database, email — eliminating the “human in the middle.”
- BullMQ + Redis: task queue — ensures ordered Agent execution under high concurrency.
- Neon (Postgres): serverless database — elastic scale, zero ops burden for a solo founder.
Roles one person carried
| Role | AI replacement | Notes |
|---|---|---|
| Architect | 30% | Ben designs the overall Agent hierarchy; AI helps refine details |
| Development | 90% | Claude Code generates nearly all code |
| Operations | 95% | AI Agents handle daily operations |
| Customer support | 95% | The support Agent responds automatically |
| Finance | 80% | Stripe integration is automated, but dispute handling exposed the blind spot |
| Oversight / audit | 0% (the problem) | Lack of human oversight led to silent failures |
3 Reproducible Key Moves
- Use specs to define each Agent’s behavioral boundary: Don’t let an AI “wing it.” Write an explicit spec.md for each Agent (role, permissions, triggers, forbidden actions, escalation rules). Use Claude conversationally to clarify requirements first, then have it generate the structured spec document — review it before implementation.
- Use the MCP protocol to connect external systems: Don’t write your own API wrappers — use an MCP Server so Agents can operate Stripe / database / email directly. This cuts implementation effort by 70%+.
- Build an “exception escalation” mechanism — don’t chase 100% automation: Polsia’s lesson is the canonical counter-example. Any step involving money, compliance, or customer complaints must have a human-review trigger. In BullMQ, set an
escalationflag on high-risk tasks; when triggered, notify a human via Slack/email.
Limits and risks
- Silent-failure risk is extremely high: 20 unhandled Stripe disputes is the most dangerous blind spot of “full automation.” Any step involving money cannot go without a human backstop.
- Dependency on frontier tooling: MCP and Claude Opus 4.6 only matured in 2025–2026; early adopters bear tooling instability risk.
- Revenue sustainability in question: $500K/month growth is unusually fast — retention rate and dispute rate need watching.
- Compliance risk: AI handling customer data and money autonomously may face compliance challenges in different jurisdictions.