What I Learned Shipping AI Agents at Scale

Deploying agents that run research and analysis in parallel taught me that the hard part isn't the model — it's the orchestration, review checkpoints, and trust boundaries around what an agent is allowed to do unsupervised.

Three lessons that generalized well beyond this one project:

1. Human review checkpoints need to be cheap, or people route around them.
2. Agents need scoped, auditable permissions — not blanket access.
3. The orchestration layer is where most of the real engineering happens.