Agents in production
What happens after the demo. Pilots that die in the long tail, the real cost of delegating work to cheaper models, and the queue that backs up once writing code stops being the bottleneck.
The delegation tax: measuring what it costs to let cheap models do the work
A pre-registered A/B between an all-cloud agent and a Claude-orchestrated local-worker hybrid. Every run passed. The meter says the real price is trust.
The bottleneck moved to code review. Don't automate the part that mattered.
Agentic tools made writing code faster, so the queue backed up at review. Pointing the same agents at the queue automates the half that never mattered.
Most agentic AI pilots die in the long tail
Why pilots flatter agentic AI and production exposes it — and what separates the few deployments that cross the gap from the many that don't.