Hundreds of agents can now write millions of lines of code. That proves output scales, not value. Value still depends on deciding, verifying and coordinating. AI amplifies whatever coordination an organisation already has. Cursor ran hundreds of agents to produce a browser in a week. Anthropic built a C compiler with parallel agents. Both were research experiments, not production software. A Google study found multi-agent setups hurt sequential tasks but help divisible ones. Cursor’s first equal-agent design slowed throughput to two or three effective agents. A hierarchy of planner and worker agents fixed this. Brooks’s Law still applies: adding agents to a late project can make it later. Cheap code makes hidden coordination failures visible faster. Review queues and unused features appear sooner.
What changed
Code was expensive and the bottleneck of software delivery.
What it unlocks
Scaling code output without proportional increases in human coordination effort.
- 1M+ lines of code in a week
- 39–70% worse sequential planning
- 80% better on divisible tasks
- $2.5B run-rate revenue for Claude Code
What you need to act on it
- a hierarchy of agents
- a single design authority
- written intent
- a strong test suite
Sources