SaaStr replaced 13 employees with 3 people + 20 AI agents. Two AI VPs cost $254/month against ~$500K of human cost. Stripped of the shock value, here are six principles for designing an agent that survives as a commercial entity.
1. Design for agents, not humans. Agents don't browse your UI, they call your API. They care about rate limits, OAuth, REST conformance, error handling, webhook reliability. Stripe scored the only A+ on SaaStr's 116-API report card: MCP server, agent toolkit for 4 frameworks, llms.txt at root, restricted keys scoped per-tool.
2. One inch wide, one mile deep. Agent ecosystems don't reward generalists. Agents repeatedly reach for tools with exceptional domain depth. Pick one capability node and own it.
3. Be the tool agents pick. Agents are a new distribution channel. Don't ask "will agents replace me" — ask "will an agent reach for me at the step where my capability is needed."
4. Draw the autonomy boundary. Agents own execution. Humans own judgment and relationships. A well-designed agent knows when to stop, when to escalate, and what it must not do. Assist agents recommend; humans decide.
5. Consistency beats brilliance. There is no set-it-and-forget-it agent. Output is B/B+ but trained daily. Design the operating loop, not just the agent.
6. The flywheel is the moat. Every call makes the agent smarter. Don't bet on the model — it resets every quarter. Avoid "feature, not a company" and "solution in search of a problem."
The agent era is a screening question: a tool agents pick, or a tool agents route around.
Top comments (0)