Friday it worked perfectly. Saturday it threw errors. Sunday, still testing, still confused. The inconsistency was maddening. Then I realized: when it succeeded, that was proof it was broken. Claude was compensating. The truth hit hard. Agents can’t spawn other agents. My entire three-tier architecture was impossible.
Three days of building taught me what I needed to know. Now I’d rebuild everything from scratch. Proper architecture this time. Twelve agents. Three tiers. Complete testing documentation. And at the end, I’d delete 6,008 lines of the old implementation without hesitation. The code had taught me what it should become.
I had a workflow that worked beautifully. Then I needed it to work somewhere else. My first instinct: copy everything and modify the duplicate. My second instinct: that’s terrible. There has to be a better way. What if the workflow didn’t know where it was running? What if it didn’t care?
I stared at an agent with twenty steps organized into four phases. It worked perfectly. But something felt wrong. If you need phases, you’re not looking at one job. You’re looking at multiple jobs pretending to be one. By the end of the day, I’d discovered a rule that would change everything: maximum ten steps.
I asked a simple question: can agents coordinate other agents? The answer was yes, and suddenly I was building seven specialists working together autonomously. But the real discovery came when the first pass missed issues. I needed it to run again. And again. Until nothing remained. Loop until clean.
For two months, I directed Claude’s every move. Read this file. Fix that bug. Update this code. It made me four times faster. Then someone mentioned “agents” and I asked what they were. The answer changed everything. What if Claude didn’t need me to orchestrate each step?