Agentic Dev
Treating coding agents as a fleet, not a chatbot
The leap in productivity isn't a smarter single agent. It's structuring many narrow agents with clear contracts, gates, and verification.
The default mental model for an AI coding tool is a very smart pair programmer you chat with. That model caps out fast. The one that scales treats agents like a fleet: many narrow workers, each with a defined job, a contract, and a verification step.
Narrow beats general
A single general agent asked to "review this PR" gives you a vague, plausible answer. Three narrow agents - one for security, one for architecture fit, one for the style and test checklist - each give you a specific, checkable answer. Narrow scope is what makes the output falsifiable.
Contracts make output composable
When an agent returns prose, you read it. When it returns structured data against a schema, you can route it, filter it, and feed it to the next stage. The difference between a demo and a pipeline is whether the handoffs are typed.
Gates keep humans in the loop where it matters
Not every step should be autonomous. The useful pattern is to let agents fan out on the cheap, reversible work - searching, drafting, finding - and insert a human gate before anything hard to undo. The agents do the breadth; you keep the judgment.
Verify, don't trust
A finding from an agent is a hypothesis, not a fact. The cheapest way to raise quality is a second agent whose only job is to try to refute the first. Plausible-but-wrong is the failure mode of these systems, and adversarial verification is the antidote.
The productivity jump isn't a better prompt. It's an architecture.