Coordinate parallel subagents for complex multi-lane tasks
Decompose complex work into parallel subagent assignments, then integrate and verify results while keeping approvals with the user.
Why it matters
Break down substantial work into independent assignments, dispatch focused subagents to execute them in parallel, then integrate and verify the combined results while maintaining user approval for consequential decisions.
Outcomes
What it gets done
Decompose complex tasks into distinct, bounded assignments with explicit outputs
Run parallel read-only scouts with scoped context for research and review lanes
Assign non-overlapping implementation work to specialized subagents with appropriate reasoning effort
Integrate outputs, resolve conflicts, and verify the final combined result
Install
Add it to your toolbox
Run in your project directory:
curl -fsSL https://spark.entire.vc/get/ag-orchestrate | bash Overview
Orchestrate
Orchestrate coordinates substantial work by breaking tasks into distinct, bounded assignments for focused subagents, then integrates their outputs into a verified result. The coordinator remains available to users and responsible for conflict resolution, verification, and the final deliverable. Use Orchestrate when a task has multiple independent research, review, or implementation lanes that benefit from parallel execution to reduce elapsed time or improve coverage. It fits situations where a coordinator must synthesize several bounded outputs into one verified result, but not for trivial tasks that should stay with the coordinator.
What it does
Orchestrate coordinates substantial work by decomposing tasks into distinct, bounded assignments for focused subagents, then integrates their outputs into a single verified result. The coordinator remains available to the user throughout and retains responsibility for the final deliverable, including conflict resolution and verification.
When to use - and when NOT to
Use Orchestrate when a task has multiple independent research, review, or implementation lanes that can run in parallel to materially reduce elapsed time or improve coverage. Use it when a coordinator must synthesize several bounded outputs into one verified result.
Do NOT use Orchestrate for trivial tasks; keep those with the coordinator. Avoid parallel agents when the task is not substantial enough to justify the added cost and coordination overhead.
Inputs and outputs
The user provides a substantial task that can be decomposed into distinct lanes. The coordinator decomposes it into bounded assignments with explicit outputs, each with all scoped context and evidence required.
The user receives an integrated, verified result after the coordinator resolves conflicts among subagent outputs and completes final verification. Approvals and externally consequential decisions remain with the user.
Workflow and technical details
The coordinator follows a six-step workflow:
- Decompose the task into distinct, bounded assignments with explicit outputs
- Run narrow, read-only scouts in parallel with low reasoning effort and no inherited conversation when the runtime supports those controls, giving each scout all scoped context and evidence
- Use medium reasoning effort for routine implementation and high reasoning effort for difficult work
- Give each subagent distinct ownership, prevent overlapping assignments, and instruct leaf workers not to delegate
- Integrate the outputs, resolve conflicts, and verify the combined result
- Keep approvals and externally consequential decisions with the user
For a repository-wide feature, assign non-overlapping agents to architecture inspection, implementation, and test review, then integrate their findings and run final verification from the coordinator. For a research brief, assign independent sources or questions to read-only scouts, reconcile disagreements, and keep final judgment with the coordinator.
Limitations
Orchestrate requires a runtime that exposes subagent or delegation tools; otherwise keep the work with the coordinator. Delegation does not authorize file mutations, public actions, purchases, or other consequential operations beyond the user's original scope. Parallel agents add cost and coordination overhead, so use them only when the task is substantial enough to benefit. The coordinator remains responsible for checking claims, changes, tests, and the final answer.
FAQ
Common questions
Discussion
Questions & comments · 0
Sign In Sign in to leave a comment.