Skills-compatible agents / Agent coordination
Dispatching Parallel Agents
Splits genuinely independent investigations into self-contained agent tasks, then reviews their combined result.
316,020 repository stars · checked 2026-11-11. Count applies to obra/superpowers, not this individual skill.
Guide by IDAKIM Digital. Source review: . Original explanation based on the linked publisher sources.
What it does
The Superpowers dispatching-parallel-agents skill is for separate problem domains that can be understood and worked on without shared state. It first asks whether failures are related; parallelism is inappropriate when one fix could explain several symptoms. Each worker receives a narrow scope, the necessary evidence, constraints and a return contract, rather than inheriting the entire parent conversation. The controller then reads the reports, checks for conflicts and verifies the integrated result. The point is controlled independent investigation, not maximizing the number of agents or treating simultaneous edits as inherently safe.
When to use it
- Investigate unrelated failures in an approval flow, batch completion and cancellation subsystem.
- Review independent documentation areas with non-overlapping file ownership.
- Compare isolated root-cause hypotheses when each investigator can use separate fixtures and resources.
Step-by-step workflow
- Group failures by likely cause and check whether a shared dependency could explain them first.
- Confirm each domain can be investigated without mutating the same files, database or runtime state.
- Write focused prompts with concrete failing cases, allowed changes, relevant context and expected reports.
- Dispatch independent tasks concurrently only when the host supports it and the user’s authorization permits delegation.
- Read every report, inspect conflicting assumptions and run the integration checks before accepting the combined changes.
Example request
Use dispatching-parallel-agents for three independently reproduced test failures. Give each worker only its failing test output, related subsystem paths and an explicit file boundary. Ask for the root cause, smallest fix and verification evidence. Do not increase arbitrary timeouts to hide races. If two failures share a fixture or service, combine that investigation instead of dispatching competing writers.
Worked example
A cancellation test fails because completion races with abort, while a static help-page test fails because an anchor changed. These domains may be independent enough for separate workers. Two transaction tests that both fail after a database reset probably are not: the fixture could be the common cause. On return, the controller checks whether either worker changed shared test helpers, even if the original scope appeared isolated. The final suite verifies the actual integrated tree, because individually passing reports cannot establish compatibility between changes.
What you get
Focused investigation reports or isolated fixes, followed by a reviewed and tested combined result.
How to evaluate the result
- Each worker has one domain, clear ownership and a concrete report contract.
- Shared fixtures and overlapping edits are detected before integration.
- The combined changes pass relevant checks after every report is inspected.
What you need
A host with concurrent-agent support, genuinely independent domains and enough evidence to write self-contained briefs. Workspace and resource isolation must be designed explicitly.
Setup and activation
Read the Superpowers README for the selected client and retain the referenced templates and supporting files. This directory reviewed source only; it did not install or execute the skill. Before adopting it, scan the exact version, inspect requested tools and permissions, and test on disposable non-sensitive inputs. A compatible instruction format does not establish identical host capabilities.
Limits and considerations
Extra agents increase coordination and context costs. Separate conversations do not isolate a shared filesystem. Popularity does not establish reliability, and worker reports remain evidence to inspect rather than authoritative proof of completion.
Related workflows
- Use systematic debugging first when the failure domains are still unknown.
- Use subagent-driven-development for an agreed implementation plan requiring sequential reviewed tasks.
License and reuse
MIT at repository level. Preserve copyright and license notices; check companion files and external dependencies for their own terms.
Further reading
Read the original source
Publisher: obra/superpowers
Source reviewed: 2026-11-11
Source reviewed; not installed, executed or security-scanned for this directory. These example requests were written by IDAKIM Digital. Upstream content can change.