Orchestrate AI coding agents in parallel

The interface is the bottleneck. We've spent years making AI agents better at coding, but almost no time making it easier to verify their work. The result? Developers spend 80% of their time reviewing and 20% prompting.

Running multiple agents at once sounds powerful until it turns into chaos: 3-4 terminals, each on a different task, and you're asking, "Which one is on auth? Did the database refactor finish?" You end up bouncing between windows, running git diff, and piecing together what changed where.

Isolation enables scale. When each agent runs in its own container with its own VS Code instance, you eliminate the confusion of shared state. Every diff is clean. Every terminal output is separate. Every verification is independent.

The issue isn't that agents aren't good — they're getting scary good. It's that our tools were designed for a different era. VS Code was built for writing code, not reviewing five parallel streams of AI-generated changes. Terminals expect sequential commands, not a fleet of autonomous workers.

Verification is non-negotiable. Code diffs are just the start. We need to see the running application, the test results, the performance metrics—all in real-time, for every agent, without switching contexts.

cmux solves this by giving each agent its own world: separate Docker container, separate VS Code, separate git state. VS Code opens with the git diff already showing. Every change is isolated to its task, so you can see exactly what each agent did — immediately — without losing context. That's what makes running 10+ agents actually workable.