| name | dispatching-parallel-agents |
| description | Use when 2+ independent tasks, failures, research lanes, exploration lanes, or path-scoped investigations can run without shared state, shared files, or sequential dependency. |
Dispatching Parallel Agents
If dispatched as a subagent for a specific task, skip this skill.
Dispatch one agent per independent problem domain. Keep shared state with the
coordinator.
Purpose: reduce wall-clock time without losing ownership boundaries.
Consumer: agents coordinating multiple test failures, validator failures,
subsystem investigations, or path-scoped implementation lanes.
Failure consequence: agents edit the same files, duplicate investigation, hide
validator debt, or return incompatible fixes.
Falsifier: one agent can resolve every failure with less context and no more
wall-clock time than parallel lanes.
Use
Use when all are true:
- 2+ failures or tasks exist.
- Each lane has distinct files, subsystem, or failure class.
- One lane's result is not needed before another starts.
- Shared files can stay with the coordinator or one named owner.
Do not use when failures likely share one root cause, need whole-system tracing,
or require the same resource/file.
Split
- By path: one test file, package, service, or doc tree per agent.
- By failure class: parser/syntax, mechanical formatting, policy contract,
runtime bug.
- By subsystem: API, worker, database, UI, deploy.
Shared files stay with the coordinator unless one agent explicitly owns them:
package.json, lockfiles, root lint config, root agent docs, CI/GitOps entry
points, generated allow-lists.
Collision Control
Before dispatch, assign each lane:
- Owned paths: files/directories the agent may edit.
- Read-only paths: context the agent may inspect but not edit.
- Artifact prefix: unique report/temp path prefix, e.g.
/tmp/agent-<lane>-<timestamp> or .agent-work/<lane>/.
- Forbidden shared paths: files owned by coordinator or another lane.
Agents must not write generic names like /tmp/report.md, /tmp/diff.patch,
notes.md, or review.md. Include lane id, task id, or timestamp in every
scratch/report path.
If two lanes need the same file, do not run them in parallel. Serialize them or
make one coordinator-owned edit after both investigations return.
Prompt Contract
Each subagent prompt must include:
- Scope: exact files, commands, logs, or subsystem.
- Goal: observable pass/fail condition.
- Constraints: allowed edits and forbidden paths.
- Unique artifact/report path prefix.
- Evidence: current failure output or reproduction command.
- Return shape: root cause, files changed, commands run with results, remaining
failures outside scope.
Do not ask a subagent to "fix all tests" or infer ownership from repository
context. Give the lane and stop rule.
Validator Lanes
Order matters when validators fail:
- Syntax/parser failures first; broken input blocks classification.
- Mechanical formatting/lint debt next; prefer formatter-safe rewrites.
- Policy failures last; fix contract or record bounded migration work.
New ignores, allow-list entries, disabled rules, or skipped validators need path,
owner, removal condition, failure consequence, and falsifier.
Integration
After agents return:
- Read every summary.
- Check file overlap and shared-state changes.
- Inspect diffs before trusting results.
- Run the narrow lane commands again if output is stale or incomplete.
- Run the integration command that proves the combined change.
Do not report completion from subagent summaries alone.
