The hardest part wasn't getting each agent to work — it was managing state between them. Agents would contradict each other or lose context mid-pipeline. I had to design explicit handoff schemas and enforce output contracts at each step. Prompting each agent with awareness of what came before it was the key shift that made the system stable.

Key learning: Agentic reliability comes from structured handoffs, not just capable models.

When generating cold emails from job descriptions, the model would invent company details or fabricate product names. Switching to a retrieval-augmented approach — grounding each generation with retrieved context — reduced factual drift significantly. The bigger lesson was learning to treat hallucination as a retrieval problem, not a prompting problem.

Key learning: Anchor generation to retrieved facts; don't rely on the model's memory.

For the horror game, I kept adding assets expecting atmosphere to follow. It didn't. The real shift came when I started removing information — limiting visibility, slowing movement, and making sound do the work that visuals couldn't. Designing constraints turned out to be more powerful than designing content. The player's imagination filled the rest.

Key learning: Immersion is an absence problem as much as a presence problem.