The Real Bottleneck with Parallel Agents Is Not Compute - It Is Git Conflicts
The Real Bottleneck with Parallel Agents Is Not Compute
I run 5 coding agents in parallel on the same repo. The bottleneck is never CPU or API rate limits. It is git conflicts.
When multiple agents work on the same codebase simultaneously, they inevitably touch the same files. Shared utilities, configuration files, type definitions, import statements - these are magnets for merge conflicts. And AI agents are terrible at resolving merge conflicts because they lack the context of what the other agent was trying to do.
The Conflict Cascade
Agent A modifies a shared utility function. Agent B adds a new import to the same file. Agent C updates the type definitions that both A and B depend on. When you try to merge all three, you get a cascade of conflicts that takes longer to resolve than the original work took to generate.
This is not a theoretical problem. It happens daily when running parallel agents at scale.
What Actually Works
Git worktrees. Each agent gets its own worktree - a separate checkout of the same repository. They work on isolated branches with isolated file systems, so there are no lock conflicts or dirty working directory issues.
The merge step still requires conflict resolution, but you control when it happens instead of having agents step on each other in real time.
The second strategy is task decomposition. Before launching parallel agents, split the work so agents operate on different parts of the codebase with minimal overlap. This requires upfront planning, but it dramatically reduces the merge pain.
The Coordination Tax
Running parallel agents is not free scaling. There is a coordination tax that grows with the number of agents. Two agents have minimal overhead. Five agents require real planning. Ten agents require a dedicated orchestration strategy.
The right number of parallel agents is the number where the coordination tax does not exceed the productivity gain. For most solo developers, that is 3-5. Beyond that, you spend more time managing agents than you save.
Fazm is an open source macOS AI agent. Open source on GitHub.