Agent Swarm Deployer

Agent Swarm Deployer

Deploy a swarm of parallel sub-agents to process massive, independent data tasks (documents, records, rows, items) and aggregate the results. Use this for data operations; use agent-army for code changes.

Contents

• references/overview.md -- swarm vs army comparison, use cases, architecture diagram
• references/swarm-design.md -- input/output schemas, batch-size and swarm-size formulas, scaling guidelines
• references/agent-brief.md -- agent brief template, data distribution methods, progress tracking
• references/aggregation-recovery.md -- merge logic, completeness validation, retry strategy, error-handling table
• references/output-formats.md -- CSV/JSON/Markdown/individual-file outputs, final summary report
• references/task-configs.md -- ready-made configs for sentiment, lead scoring, content generation, summarization

Workflow

1. Understand the task. Pin down five things before deploying anything: data source, operation per item, output format, output destination, and quality/validation requirements. If any is ambiguous, ask the user first -- a wrong spec wastes all agent compute.

2. Intake and inventory. Glob/Bash to locate and count items. Read 3-5 samples to learn structure. Estimate tokens per item and total. Report an intake summary (source, total count, item format, sample structure, token estimate).

3. Detect input schema and define output schema. Derive the input schema from samples; define the exact output schema the task requires. See references/swarm-design.md.

4. Design the swarm. Compute batch size from token budget (70% of ~200K usable context per agent) and swarm size from total items. Cap at 20 agents per wave; split into waves beyond that. Present the swarm plan and agent assignments, then get approval. See references/swarm-design.md.

5. Prepare agent briefs. Build a self-contained brief per agent: role, task, input data, output schema with example, quality rules, error protocol, and strict JSON output format. See references/agent-brief.md.

6. Distribute data and deploy. Choose a distribution method for the source type (pre-split CSVs/JSON with Bash; embed inline for small sets; pass file paths for directories). Launch up to 20 agents in parallel via the Agent tool with run_in_background: true, sending all calls in one message. Run later waves after the prior wave completes. See references/agent-brief.md.

7. Track progress. As agents return, record status, processed counts, and cumulative coverage. See references/agent-brief.md.

8. Collect and aggregate. Parse each agent's JSON; validate schema, completeness, and duplicates. Merge into one ordered output and extract failures. Report an aggregation summary with a coverage check and failure analysis. See references/aggregation-recovery.md.

9. Recover failures. Queue all failed and skipped items, deploy a retry agent with enhanced instructions, cap at 2 retries, and mark survivors "unrecoverable". Flag the user if unrecoverable items exceed 10%. See references/aggregation-recovery.md.

10. Write output and summarize. Produce the requested format (CSV, JSON, Markdown, or individual files) plus a final summary covering execution, results, quality metrics, patterns observed, and cost. See references/output-formats.md.

Anti-Patterns to Avoid

1. Do not use a swarm for sequential tasks. If item N depends on item N-1, use a chain instead. 2. Do not deploy one agent per item. Batch items; one-per-agent wastes overhead. 3. Do not skip schema definition. Without a schema, merging results from many agents becomes unreliable. 4. Do not ignore failures. At 99% success, 1% of 10,000 items is still 100 failures. Always run retries. 5. Do not deploy without a sample run. Process 5 items manually first to validate the task and output quality before scaling.

Agent Brief Template and Data Distribution

Brief Requirements

Each swarm agent receives a self-contained brief that includes:

1. Role: "You are a data processing agent in a swarm of 25. Your job is to process items 51-100." 2. Task description: exactly what to do with each item (the operation). 3. Input data: the actual data items for this batch (embedded in the brief or read from a file range). 4. Output schema: the exact format for results, with a concrete example. 5. Quality rules: validation criteria, edge case handling, what to do with malformed items. 6. Error protocol: "If an item cannot be processed, mark it as failed with a reason. Do not skip items silently." 7. Output format: "Return your results as a JSON array. Each element must match the output schema."

Brief Template

You are swarm-agent-{N}, processing batch {N} of {TOTAL_BATCHES} in a data processing swarm.

## Your Task
{TASK_DESCRIPTION}

## Input Data
You will process the following {BATCH_SIZE} items:

{ITEMS_AS_STRUCTURED_DATA}

## Output Schema
For each item, produce a result matching this schema:
{OUTPUT_SCHEMA_WITH_EXAMPLE}

## Example
Input:
{EXAMPLE_INPUT}

Expected output:
{EXAMPLE_OUTPUT}

## Quality Rules
- {RULE_1}
- {RULE_2}
- {RULE_3}
- Every item MUST appear in your output, even if processing failed.
- For failed items, set processing_status to "failed" and include an error_reason field.

## Error Handling
- Malformed input: Mark as "failed", reason: "malformed input: {description}"
- Ambiguous data: Make your best judgment, set confidence to < 0.5
- Missing required field: Mark as "skipped", reason: "missing {field}"

## Output Format
Return a JSON object with this structure:
{
  "agentId": "swarm-agent-{N}",
  "batchRange": "{START}-{END}",
  "totalProcessed": {NUMBER},
  "totalSuccess": {NUMBER},
  "totalFailed": {NUMBER},
  "totalSkipped": {NUMBER},
  "results": [
    {OUTPUT_SCHEMA_ITEM_1},
    {OUTPUT_SCHEMA_ITEM_2},
    ...
  ],
  "errors": [
    {"itemIndex": N, "reason": "description"}
  ],
  "notes": "Any observations about the data quality or patterns"
}

Data Distribution

Choose a distribution method based on the data source:

For CSVs and large files, pre-split before deploying:

# Split a CSV into batches of 50 rows (preserving header)
head -1 data.csv > header.csv
tail -n +2 data.csv | split -l 50 - batch_
for f in batch_*; do cat header.csv "$f" > "batches/${f}.csv" && rm "$f"; done
rm header.csv

Progress Tracking

As agents complete, report progress:

## Swarm Progress

Wave 1 (20 agents):
[################----] 16/20 complete

| Agent | Status | Processed | Success | Failed | Duration |
|-------|--------|-----------|---------|--------|----------|
| swarm-01 | DONE | 50 | 48 | 2 | 45s |
| swarm-02 | DONE | 50 | 50 | 0 | 38s |
| swarm-03 | RUNNING | -- | -- | -- | -- |
| ... | ... | ... | ... | ... | ... |

Cumulative: 800/1247 items processed (64%)

Aggregation, Recovery, and Error Handling

Gather Agent Outputs

As each agent completes, collect its JSON output, then parse and validate:

1. Schema validation: does each result match the output schema? 2. Completeness check: does the agent's result count match its batch size? 3. Duplicate detection: check for duplicate item IDs across agents. 4. Error extraction: pull out all failed/skipped items for the retry queue.

Merge Results

Combine all agent results into a single unified output:

# Conceptual merge logic
merged_results = []
failed_items = []
skipped_items = []

for agent_output in all_agent_outputs:
    for result in agent_output["results"]:
        if result["processing_status"] == "success":
            merged_results.append(result)
        elif result["processing_status"] == "failed":
            failed_items.append(result)
        elif result["processing_status"] == "skipped":
            skipped_items.append(result)

# Sort by original item order
merged_results.sort(key=lambda x: x["original_index"])

Validate Completeness

## Aggregation Report

- Total items in input: 1,247
- Total results received: 1,247
- Successful: 1,198 (96.1%)
- Failed: 34 (2.7%)
- Skipped: 15 (1.2%)

### Coverage Check
- Items with results: 1,247 / 1,247 (100% coverage)
- Missing items: 0
- Duplicate results: 0

### Failure Analysis
| Failure Reason | Count |
|---------------|-------|
| Malformed input | 12 |
| Missing required field | 8 |
| Ambiguous data | 14 |

Proceed to retry failed items? (Y / skip / manual review)

Error Recovery

Retry Queue

Collect all failed and skipped items into a retry batch:

Retry batch: 49 items (34 failed + 15 skipped)
Retry strategy: Single agent with enhanced instructions

Enhanced Retry Brief

You are the retry agent. These items failed or were skipped in the first pass.
For each item, I am providing the original item AND the failure reason from the first attempt.

Your job:
1. Try harder -- use more creative interpretation for ambiguous items
2. For truly malformed items, extract whatever you can and note what is missing
3. For items that failed due to missing fields, infer the field if possible or mark as "unrecoverable"

The bar is lower for retries: partial results are better than no results.

Retry Limits

• Maximum retries: 2 (original attempt + 2 retries = 3 total attempts)
• After max retries: mark item as "unrecoverable" and include in the final report
• Unrecoverable threshold: if > 10% of items are unrecoverable, flag to the user for manual review

Error Handling Reference

Output Formats and Final Summary

Write the final output in the format the user requested.

CSV Output

# Write header + all successful results as CSV
# Include a separate failures.csv for failed items

JSON Output

{
  "metadata": {
    "task": "description of what was processed",
    "totalItems": 1247,
    "successfulItems": 1210,
    "failedItems": 22,
    "unrecoverableItems": 15,
    "processingDate": "2026-04-10T12:00:00Z",
    "swarmSize": 25,
    "waves": 2
  },
  "results": [...],
  "failures": [...],
  "summary": {
    "key_patterns": "...",
    "notable_findings": "...",
    "data_quality_notes": "..."
  }
}

Markdown Report

# Processing Results: [Task Name]

## Summary
- Processed: 1,247 items
- Success rate: 97%
- Key findings: [aggregated insights]

## Results Table
| Item | Result Field 1 | Result Field 2 | Status |
|------|---------------|----------------|--------|
| ... | ... | ... | ... |

## Failures
| Item | Reason | Attempted Retries |
|------|--------|-------------------|
| ... | ... | ... |

Individual Files

For tasks where each item produces a standalone document (e.g., generating 500 blog post outlines):

output/
  001-item-name.md
  002-item-name.md
  ...
  500-item-name.md
  _summary.md
  _failures.md

Final Summary Report

Always end with a comprehensive summary:

## Swarm Processing Complete

### Execution Summary
- Task: [description]
- Data source: [source]
- Total items: 1,247
- Swarm size: 25 agents across 2 waves
- Total processing time: ~3 minutes

### Results
- Successful: 1,210 (97.0%)
- Failed (recovered on retry): 22 (1.8%)
- Unrecoverable: 15 (1.2%)
- Output written to: [path]

### Quality Metrics
- Schema compliance: 100% of successful results match output schema
- Average confidence: 0.82
- Items flagged for review: 37 (low confidence < 0.5)

### Patterns Observed
- [Any interesting patterns the swarm noticed across the data]
- [Data quality issues found]
- [Recommendations for future processing]

### Cost
- Agents deployed: 25 (Wave 1) + 1 (retry) = 26
- Estimated tokens consumed: ~1.2M input + ~400K output

Overview: Swarm vs Army, Use Cases, Architecture

Agent Army vs Agent Swarm

Use Cases

Document Processing

• Analyze 500 customer support tickets for sentiment and categorization
• Extract key information from 200 legal contracts
• Summarize 1000 research paper abstracts
• Parse 300 resumes for qualification matching

Dataset Analysis

• Score and rank 2000 leads by ICP fit
• Classify 5000 product reviews by topic and sentiment
• Audit 1000 blog posts for SEO compliance
• Grade 500 sales call transcripts against a methodology

Bulk Content Generation

• Generate personalized email first lines for 500 prospects
• Create product descriptions for 1000 SKUs
• Write social media posts for 200 blog articles
• Generate meta descriptions for 800 web pages

Data Transformation

• Convert 1000 CSV rows into structured JSON objects
• Normalize 500 address records
• Translate 300 support articles into 5 languages
• Reformat 2000 database records from schema A to schema B

Architecture

Commander (you)
 |
 |-- Phase 1: Intake & Inventory
 |    |-- Count total items
 |    |-- Sample items for schema detection
 |    |-- Estimate token budget per item
 |
 |-- Phase 2: Swarm Design
 |    |-- Calculate optimal batch size
 |    |-- Determine number of swarm agents
 |    |-- Define result schema
 |
 |-- Phase 3: Deploy Swarm (parallel)
 |    |-- Agent S1: items 1-50      ---\
 |    |-- Agent S2: items 51-100     ---|
 |    |-- Agent S3: items 101-150    ---|-- All run in parallel
 |    |-- Agent S4: items 151-200    ---|
 |    |-- ...                       ---/
 |
 |-- Phase 4: Collect & Aggregate
 |    |-- Gather all agent results
 |    |-- Merge into unified output
 |    |-- Identify failures
 |
 |-- Phase 5: Recovery (if needed)
 |    |-- Retry failed items
 |    |-- Fill gaps
 |
 |-- Phase 6: Deliver
 |    |-- Write final output
 |    |-- Generate summary report

Swarm Design: Schemas, Batch Sizing, Swarm Sizing

Schema Detection and Definition

Input Schema (from sampled items)

{
  "inputSchema": {
    "type": "object",
    "fields": {
      "name": "string",
      "email": "string",
      "company": "string",
      "title": "string",
      "linkedin_url": "string"
    }
  }
}

Output Schema (defined by the task)

{
  "outputSchema": {
    "type": "object",
    "fields": {
      "name": "string (from input)",
      "company": "string (from input)",
      "personalized_first_line": "string (generated, 1-2 sentences)",
      "pain_point_guess": "string (inferred from title + company)",
      "confidence": "number (0-1)",
      "processing_status": "enum: success | failed | skipped"
    }
  }
}

Calculate Batch Size

The batch size determines how many items each agent processes.

Formula:

usable_tokens = 200,000 * 0.70 = 140,000
tokens_per_item = input_tokens + output_tokens + 100 (overhead)
batch_size = floor(usable_tokens / tokens_per_item)

Practical limits:

• Minimum batch size: 5 items (agent overhead is not worth it for fewer)
• Maximum batch size: 200 items (beyond this, agent context gets cluttered)
• Sweet spot: 20-80 items per agent for most text processing tasks

Determine Swarm Size

total_items = 1,247
batch_size = 50
swarm_size = ceil(1,247 / 50) = 25 agents

Practical limits:

• Maximum parallel agents: 20 per wave (prevents system overload)
• If swarm_size > 20: deploy in waves. Wave 1: agents 1-20. Wave 2: agents 21-25.

Present Swarm Plan

## Swarm Plan

- Total items: 1,247
- Batch size: 50 items per agent
- Swarm size: 25 agents
- Waves: 2 (Wave 1: 20 agents, Wave 2: 5 agents)
- Estimated processing: All items covered
- Output format: Single CSV file with all results

### Agent Assignments
| Agent | Items | Range | Notes |
|-------|-------|-------|-------|
| swarm-01 | 50 | items 1-50 | -- |
| swarm-02 | 50 | items 51-100 | -- |
| ... | ... | ... | -- |
| swarm-25 | 47 | items 1201-1247 | Partial batch |

Proceed? (Y to deploy / N to adjust batch size)

Scaling Guidelines

Swarm Configurations for Common Tasks

Sentiment Analysis (1000 reviews)

task: Classify each review as positive/negative/neutral with confidence score
batch_size: 100  # Reviews are short, pack more per agent
swarm_size: 10
output_schema:
  review_id: string
  sentiment: enum(positive, negative, neutral, mixed)
  confidence: float(0-1)
  key_phrases: string[]
  summary: string(1 sentence)

Lead Scoring (2000 contacts)

task: Score each lead 1-100 based on ICP fit criteria
batch_size: 40  # Each lead needs more analysis context
swarm_size: 50
waves: 3
output_schema:
  lead_id: string
  score: int(1-100)
  icp_match: object
    company_size: bool
    industry: bool
    tech_stack: bool
    title_seniority: bool
  buying_signals: string[]
  recommended_action: enum(hot, warm, nurture, disqualify)

Content Generation (500 product descriptions)

task: Write a 100-word product description from product data
batch_size: 25  # Output is longer, needs more generation tokens
swarm_size: 20
output_schema:
  product_id: string
  title: string
  description: string(100 words)
  key_features: string[3]
  seo_keywords: string[5]

Document Summarization (300 papers)

task: Summarize each paper in 3 bullet points with key findings
batch_size: 15  # Papers are long, fewer per agent
swarm_size: 20
output_schema:
  paper_id: string
  title: string
  summary_bullets: string[3]
  key_finding: string
  methodology: string
  relevance_score: float(0-1)