AI agents are often described as “having 100 junior engineers at your disposal.” That analogy is useful & also a warning.

If you give 100 juniors shared state, non-hermetic builds, and no guardrails, you don’t get leverage.

You get Gas Town: noisy, resource-starved, everyone tripping over each other to ship something. This is where many multi-agent systems break down.

A common mistake is confusing coordination with isolation. Git worktrees, shared repos, and shared caches help humans move fast, but they are not sandbox boundaries. They isolate files, not background processes, build daemons, ports, CPU/memory, or flaky global caches.

In real codebases (Gradle, Android, large JS monorepos), this guarantees interference and agent drift.

That’s how you end up with agents fighting over ports and caches like scavengers over fuel - locally “working,” globally unstable. Gas Town dynamics.

The fix isn’t “smarter agents,” it’s harder execution boundaries. Agents should run in bounded, disposable sandboxes with explicit CPU/mem/time budgets, isolated caches and ports, and reproducible toolchains. The contract should be boring and strict: checkout → make X change → build/test → report artifacts.

Another failure mode is agents orchestrating other agents without governance. Unbounded fan-out creates feedback loops, wasted compute, and non-deterministic outcomes. That’s not an organization - that’s Gas Town scaling laws. Multi-agent systems need limits: caps on depth and concurrency, budgets, circuit breakers, capability-based permissions.

There’s also a hardware reality check. Not everyone has a $5k laptop, and not every repo can run 10 builds locally. Serious agent systems must support disposable sandboxes on remote compute (SSH, k8s, VMs), not just local execution.

Finally, auditability matters. It’s not enough to know what happened. You need to know what was tried, how to reproduce it, what evidence was produced, and why alternatives were rejected. That’s how agents become trustworthy collaborators instead of noisy automation.

Agents aren’t infallible. Containers aren’t magic. Worktrees aren’t sandboxes.

Without bounded execution & governance, you don’t get a team - you get Gas Town.

With them, you get scale.

Also wrote more about Containers Sandboxes for AI agents here.