subagent-driven-analysis

Overview

Execute an approved experiment plan by dispatching a fresh subagent per analysis task, with two-stage review after each: (1) does the analysis match the experiment design? (2) are the results scientifically sound?

Core principle: Fresh subagent per task + two-stage review (design compliance then scientific soundness) = rigorous, reproducible results.

When to Use

digraph when_to_use {
    "Have experiment design?" [shape=diamond];
    "Tasks mostly independent?" [shape=diamond];
    "subagent-driven-analysis" [shape=box];
    "Re-plan or use manual execution" [shape=box];

    "Have experiment design?" -> "Tasks mostly independent?" [label="yes"];
    "Have experiment design?" -> "Re-plan or use manual execution" [label="no — use experimental-design first"];
    "Tasks mostly independent?" -> "subagent-driven-analysis" [label="yes"];
    "Tasks mostly independent?" -> "Re-plan or use manual execution" [label="no — tightly coupled"];
}

The Process

digraph process {
    rankdir=TB;

    subgraph cluster_per_task {
        label="Per Analysis Task";
        "Dispatch analysis subagent with full task spec" [shape=box];
        "Subagent asks questions?" [shape=diamond];
        "Answer questions, provide context" [shape=box];
        "Subagent runs analysis, logs parameters, saves results" [shape=box];
        "Dispatch design-compliance reviewer" [shape=box];
        "Analysis matches experiment design?" [shape=diamond];
        "Subagent fixes design gaps" [shape=box];
        "Dispatch scientific-soundness reviewer" [shape=box];
        "Results scientifically sound?" [shape=diamond];
        "Subagent addresses scientific issues" [shape=box];
        "Mark task complete" [shape=box];
    }

    "Read experiment design, extract all tasks" [shape=box];
    "More tasks remain?" [shape=diamond];
    "Dispatch final results reviewer" [shape=box];
    "Invoke validation-driven-research" [shape=doublecircle];

    "Read experiment design, extract all tasks" -> "Dispatch analysis subagent with full task spec";
    "Dispatch analysis subagent with full task spec" -> "Subagent asks questions?";
    "Subagent asks questions?" -> "Answer questions, provide context" [label="yes"];
    "Answer questions, provide context" -> "Dispatch analysis subagent with full task spec";
    "Subagent asks questions?" -> "Subagent runs analysis, logs parameters, saves results" [label="no"];
    "Subagent runs analysis, logs parameters, saves results" -> "Dispatch design-compliance reviewer";
    "Dispatch design-compliance reviewer" -> "Analysis matches experiment design?";
    "Analysis matches experiment design?" -> "Subagent fixes design gaps" [label="no"];
    "Subagent fixes design gaps" -> "Dispatch design-compliance reviewer" [label="re-review"];
    "Analysis matches experiment design?" -> "Dispatch scientific-soundness reviewer" [label="yes"];
    "Dispatch scientific-soundness reviewer" -> "Results scientifically sound?";
    "Results scientifically sound?" -> "Subagent addresses scientific issues" [label="no"];
    "Subagent addresses scientific issues" -> "Dispatch scientific-soundness reviewer" [label="re-review"];
    "Results scientifically sound?" -> "Mark task complete" [label="yes"];
    "Mark task complete" -> "More tasks remain?";
    "More tasks remain?" -> "Dispatch analysis subagent with full task spec" [label="yes"];
    "More tasks remain?" -> "Dispatch final results reviewer" [label="no"];
    "Dispatch final results reviewer" -> "Invoke validation-driven-research";
}

Analysis Subagent Prompt Template

When dispatching a subagent for an analysis task, include:

## Context
You are executing step [N] of a computational biology experiment.

**Overall hypothesis:** [paste hypothesis]
**This task:** [paste full task spec from experiment design]
**Previous results:** [summary of prior steps if relevant]

## Requirements
1. Follow the experiment design EXACTLY — use specified tools, parameters, data sources
2. Log ALL parameters used (even defaults)
3. Record software versions
4. Save all intermediate and final outputs with clear filenames
5. Set random seed to [specified seed] for reproducibility
6. Run the specified control analyses alongside the main analysis
7. Do NOT interpret results — just produce the data. Interpretation comes later.

## Output Format
Save results to: `results/step-[N]-[description]/`
Include: raw output, parsed summary, parameter log, runtime log

Design Compliance Review Checklist

The design-compliance reviewer checks:

• [ ] Correct tool and version used?
• [ ] All parameters match the experiment design?
• [ ] Correct data sources and accessions?
• [ ] Controls were run?
• [ ] Random seed was set as specified?
• [ ] Outputs saved in the correct format and location?
• [ ] No extra analyses added (not in the design)?
• [ ] No analyses skipped?

Scientific Soundness Review Checklist

The scientific-soundness reviewer checks:

• [ ] Results are numerically plausible? (e.g., e-values in expected range)
• [ ] No obvious errors? (empty results, NaN values, corrupted output)
• [ ] Controls produced expected results? (positive control positive, negative control negative)
• [ ] Data formats are correct and parseable?
• [ ] Results are reproducible? (could be re-run from logged parameters)

Red Flags — STOP

Never:

• Start analysis before experiment design is approved
• Skip either review stage (design compliance OR scientific soundness)
• Interpret results during analysis (interpretation comes in results-review)
• Run extra analyses not in the design ("while we're at it...")
• Ignore subagent questions
• Accept "close enough" on design compliance
• Dispatch multiple analysis subagents in parallel on shared data (conflicts)
• Let the analysis subagent modify input data

If reviewer finds issues:

• Same subagent fixes them
• Reviewer reviews again
• Repeat until approved
• Never skip re-review

Advantages

• Fresh context per task — no confusion from prior analyses
• Two-stage review ensures both protocol adherence and scientific validity
• Reproducibility built in — every parameter logged by design
• No interpretation bias — analysis and interpretation are separated

After All Tasks

• Run final results reviewer across all completed analyses
• Check for consistency across steps
• Invoke validation-driven-research for statistical validation