From pair programming to fleet management

The bottleneck in AI-assisted development is often not the model, it is the cost of switching context. Every time an engineer stops to set up a task, wait for a result and then reload the problem in their head, the speed gain from the model leaks away. One assistant in a chat pane is still pair programming, and pair programming does not scale past one pair.

Multi-agent orchestration in the editor changes the shape of the work. A local agent handles strategy and quick answers. Background agents work in isolated worktrees on well-scoped tasks. Cloud agents take the heavy lifting, like large refactors or long test runs. All of them draw on shared context, so nobody has to re-explain the codebase each time.

The useful distinction for leaders is between automation and orchestration. Automation is doing things right: one agent completing one task well. Orchestration is doing the right things: deciding which work goes to which agent, in what order, with what checks before anything merges. The second is an architecture and management problem, and that is where I see most teams underinvesting. The payoff is asynchronous execution. Engineers stop babysitting a single session and start reviewing the output of several, which is a different skill and needs different guardrails.

Build the factory first; the widgets will follow.