The original title is: "I Tried to Design an Entire AI Software Testing Company. Here's the Architecture I'd Actually Build."

Short summary

Ambitious AI testing organization (30 agents) redesigned for practical shipping: build the artifact graph and approval state machine first, then scale agents. V1 scope is API testing—deterministic, verifiable, winnable. Expand to UI/mobile from a position of strength.

• •Focus on the durable core (artifact graph + approval lifecycle) before agents; agent specs are cheap but reliable engineering is expensive.
• •Start with API testing v1—deterministic, verifiable, survivable last-mile. UI/mobile is v2-v3, after proving the concept.
• •Route models by task class to optimize LLM costs; avoid the trap of thousands of calls and $40+ per upload.

Generated with AI, which can make mistakes.