Parallelism helps when tasks are independent and read-heavy. It hurts when several agents write overlapping files or require constant shared decisions.
| Mechanism | Best fit |
|---|---|
| Subagents | Parallel exploration, testing, triage, or specialist review |
| Cloud task | Hosted asynchronous work against a pushed repository |
| Local worktree | Isolated work using your machine and local tools |
Spawn three read-only subagents:
1. Trace the authentication flow.
2. Identify test gaps around token refresh.
3. Review logging for credential exposure.
Wait for all three. Return one deduplicated report
with evidence and file references. Do not edit files.Subagents should return distilled findings, not raw logs. Keep requirements, decisions, and synthesis in the main thread.
Two tasks are safe to parallelize when neither requires the other’s intermediate decisions, they do not edit overlapping state, and their outputs can be merged by a clear rule.
| Shape | Parallel? |
|---|---|
| Security review and test-gap review of the same diff | Yes; read-only perspectives |
| Two agents redesigning the same module | No; shared decisions and files |
| Compare two prototypes in separate worktrees | Yes; isolate and define comparison criteria |
| Schema change followed by dependent API changes | Usually sequential |
Each subagent performs its own model and tool work. Parallelism costs more tokens and adds synthesis overhead. Use it when elapsed time, independent evidence, or context isolation is worth that cost.
A cloud task starts from repository state available to its environment, not every untracked local artifact. Confirm the commit, setup script, dependencies, secrets policy, and internet settings before delegating.
Take one review question and split it into two independent evidence-gathering tasks. Define authority, output schema, evidence requirements, and the rule the main agent should use to deduplicate or resolve conflicting findings.