dispatching-parallel-agents

Dispatch one agent per independent problem domain with a focused, self-contained prompt specifying scope, goal, constraints, and expected output. Each agent gets isolated context constructed by you, never your session history, so they work concurrently without interference.

Use parallel dispatch when failures are independent, such as multiple test files failing with different root causes or separate subsystems broken without shared state. Avoid it when failures are related, when fixes might overlap, or when agents would write to the same files.

Partition work by write-set, not by task list: trace each fix to its actual edit sites and assign each file to exactly one agent. Reserve shared aggregator files for the dispatcher, and commit with explicit pathspecs after re-verifying branch and index state.

Include a specific scope, clear goal, constraints on what not to change, and the expected return format. Provenance-stamp any measured facts with source paths and dates, label unverified claims, and state out-of-scope blockers as STOP-and-report conditions rather than silent route-arounds.

Read-only evaluation fan-outs skip collision safeguards but need a dual output contract: each agent writes full detail to its own scratch file and returns a compact summary under fixed headings. This keeps findings comparable and prevents truncation or loss during synthesis.

Agents diverge when the prompt pins only the core edit but not its edges, such as whether to update sibling references to a renamed value. State the follow-on scope explicitly in one sentence so every agent produces coherent, uniform output.