Dispatching Parallel Agents

Dispatching Parallel Agents

Overview

When you have multiple independent experiments (different datasets, different methods, different baselines), running them sequentially wastes time. Each experiment is independent and can happen in parallel.

Core principle: Dispatch one agent per independent problem domain. Let them work concurrently.

Use this skill alongside:

• amplify:experiment-execution for Phase 4 orchestration
• amplify:results-verification-protocol before claiming completion
• amplify:results-integration when consolidating outputs

When to Use

digraph when_to_use {
    "Multiple experiments?" [shape=diamond];
    "Are they independent?" [shape=diamond];
    "Single agent runs all" [shape=box];
    "One agent per experiment" [shape=box];
    "Can they work in parallel?" [shape=diamond];
    "Sequential agents" [shape=box];
    "Parallel dispatch" [shape=box];

    "Multiple experiments?" -> "Are they independent?" [label="yes"];
    "Are they independent?" -> "Single agent runs all" [label="no - related"];
    "Are they independent?" -> "Can they work in parallel?" [label="yes"];
    "Can they work in parallel?" -> "Parallel dispatch" [label="yes"];
    "Can they work in parallel?" -> "Sequential agents" [label="no - shared state"];
}

Use when:

• 3 independent experiments on different datasets
• Multiple baseline implementations to evaluate
• Each experiment can be understood without context from others
• No shared state between experiments

Don't use when:

• Experiments are related (result of one informs another)
• Need to understand full system state
• Agents would interfere with each other

The Pattern

1. Identify Independent Domains

Group experiments by what they investigate:

• Experiment A: Baseline model on dataset X
• Experiment B: Proposed method on dataset Y
• Experiment C: Ablation study on component Z

Each domain is independent - running the baseline doesn't affect the ablation study.

2. Create Focused Agent Tasks

Each agent gets:

• Specific scope: One experiment or method
• Clear goal: Run this experiment and collect results
• Constraints: Don't change shared code or other experiments
• Expected output: Summary of findings, metrics, and artifacts

3. Dispatch in Parallel

Use the Task tool to launch subagents. The Task tool is the standard mechanism for spawning independent subagents in Cursor and similar AI development environments.

How to invoke: Call multiple Task tools in a single response. Each Task runs as an independent subagent with its own context.

Call Task tool with:
  description: "Run baseline on dataset X"
  prompt: |
    [Full self-contained prompt — see Agent Prompt Structure below]
  subagent_type: "generalPurpose"

Call Task tool with:
  description: "Run proposed method on dataset Y"
  prompt: |
    [Full self-contained prompt]
  subagent_type: "generalPurpose"

Call Task tool with:
  description: "Run ablation on component Z"
  prompt: |
    [Full self-contained prompt]
  subagent_type: "generalPurpose"

Critical: All three Task calls go in the SAME message, so they execute concurrently. If you put them in separate messages, they run sequentially.

4. Review and Integrate

When agents return:

• Read each summary
• Verify results don't conflict
• Run validation checks
• Integrate all findings

Agent Prompt Structure

Good agent prompts are:

1. Focused - One clear experiment or problem domain
2. Self-contained - All context the subagent needs (it has NO access to your conversation history)
3. Specific about output - What should the agent return?
4. Include file paths - The subagent starts in the workspace root; give absolute or relative paths

Call Task tool with:
  description: "Run baseline on sentiment dataset"
  prompt: |
    Run the baseline experiment on the sentiment analysis dataset.

    Project root: [workspace path]

    Steps:
    1. Load the pre-trained model from checkpoints/baseline-v2
    2. Evaluate on test split of sentiment-benchmark
    3. Collect metrics: accuracy, F1, precision, recall

    Configuration:
    - Batch size: 32
    - Use GPU if available
    - Log results to experiments/baseline-sentiment/

    Do NOT modify the model code or training pipeline.

    Return: Summary of metrics, any anomalies observed, and path to saved results.
  subagent_type: "generalPurpose"

Common Mistakes

Bad: Too broad: "Run all the experiments" - agent gets lost Good: Specific: "Run baseline on sentiment dataset" - focused scope

Bad: No context: "Evaluate the model" - agent doesn't know which model or data Good: Context: Specify model path, dataset, metrics, and configuration

Bad: No constraints: Agent might refactor everything Good: Constraints: "Do NOT change model code" or "Evaluation only"

Bad: Vague output: "Run it" - you don't know what was measured Good: Specific: "Return summary of metrics and path to artifacts"

When NOT to Use

Related experiments: Result of one informs the next - run sequentially Need full context: Understanding requires seeing entire research pipeline Exploratory analysis: You don't know what to investigate yet Shared state: Agents would interfere (writing same output files, using same GPU)

Real Example from Session

Scenario: 3 independent experiments needed after method development

Experiments:

• Dataset A evaluation: Baseline model performance (accuracy, F1)
• Dataset B evaluation: Proposed method with hyperparameter sweep
• Ablation study: Component contribution analysis on validation set

Decision: Independent domains - each dataset/method combination is separate

Dispatch (all three Task calls in one message):

Task(description="Baseline on Dataset A", prompt="...", subagent_type="generalPurpose")
Task(description="Proposed method on Dataset B", prompt="...", subagent_type="generalPurpose")
Task(description="Ablation study", prompt="...", subagent_type="generalPurpose")

Results:

• Agent 1: Baseline accuracy 78.3%, F1 0.76 - logged to experiments/baseline-A/
• Agent 2: Best config found (lr=3e-4, layers=6), accuracy 84.1% - logged to experiments/proposed-B/
• Agent 3: Attention component contributes +3.2%, embedding component +1.8% - logged to experiments/ablation/

Integration: All results independent, no conflicts, comprehensive comparison table built

Time saved: 3 experiments completed in parallel vs sequentially

Key Benefits

1. Parallelization - Multiple experiments happen simultaneously
2. Focus - Each agent has narrow scope, less context to track
3. Independence - Agents don't interfere with each other
4. Speed - 3 experiments completed in time of 1

Verification

After agents return:

1. Review each summary - Understand what was measured and found
2. Check for conflicts - Did agents write to the same output paths?
3. Run validation - Verify all results are reproducible
4. Spot check - Agents can make systematic errors

Before declaring success, invoke:

• amplify:results-verification-protocol
• amplify:claim-evidence-alignment (if claims are being prepared for reporting/writing)

Real-World Impact

From research session:

• 3 independent experiments needed
• 3 agents dispatched in parallel
• All experiments completed concurrently
• All results integrated successfully
• Zero conflicts between agent outputs