Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.sirenspec.dev/llms.txt

Use this file to discover all available pages before exploring further.

A professor uploads N student papers. Each paper gets its own swrm: three specialist agents run concurrently (editor, AI detector, grader), and a synthesis step folds their verdicts into a single grade report. The factory collects all N reports; a gradebook compiler assembles the final markdown table. This is the swrm factory pattern — every for_each item spawns a complete mini-swrm, not just a single agent.

​
What it demonstrates

  • FactoryNode.swrm — each item in for_each spawns a full swrm instead of a single agent
  • {{ item }} / {{ index }} / {{ total }} inside swrm agent prompts and the synthesis prompt
  • Per-item synthesis: each paper gets its own grade report before results are collected
  • on_failure: continue — a single bad paper does not abort the rest of the batch
  • concurrency: 3 — at most 3 papers graded simultaneously, each paper’s 3 agents run fully in parallel

​
Run it

sirenspec run docs/cookbook/grading-factory/workflow.yaml
# Pass your own papers as a JSON array:
sirenspec run docs/cookbook/grading-factory/workflow.yaml \
  --input '["Paper one text here.", "Paper two text here."]'

​
Workflow

docs/cookbook/grading-factory/workflow.yaml
version: "0.1"

agents:
  gradebook_compiler:
    model: "anthropic:claude-haiku-4-5-20251001"
    system: |
      Compile the individual grade reports into a markdown gradebook.
      Reports: {{ grade_papers.output }}

nodes:
  grade_papers:
    type: factory
    for_each: "{{ inputs.message }}"
    concurrency: 3
    on_failure: continue
    writes: working.grade_reports
    swrm:
      concurrency: 3
      agents:
        - id: editor
          provider: anthropic
          model: claude-haiku-4-5-20251001
          prompt: "Review paper {{ index }} of {{ total }}: {{ item }}"
        - id: ai_detector
          provider: openai
          model: gpt-4o-mini
          prompt: "Assess AI likelihood for paper {{ index }} of {{ total }}: {{ item }}"
        - id: grader
          provider: anthropic
          model: claude-haiku-4-5-20251001
          prompt: "Grade paper {{ index }} of {{ total }}: {{ item }}"
      synthesis:
        provider: anthropic
        model: claude-haiku-4-5-20251001
        prompt: |
          Produce a grade report for paper {{ index }} of {{ total }}.
          Editor: {{ grade_papers.agents.editor.output }}
          AI check: {{ grade_papers.agents.ai_detector.output }}
          Grade: {{ grade_papers.agents.grader.output }}

  compile_gradebook:
    agent: gradebook_compiler
    writes: output.gradebook

edges:
  - from: grade_papers
    to: compile_gradebook

​
Graph

​
How the swrm factory works

For a batch of 3 papers, the executor runs: 
Paper 0 → [editor, ai_detector, grader] → synthesis → grade report 0
Paper 1 → [editor, ai_detector, grader] → synthesis → grade report 1
Paper 2 → [editor, ai_detector, grader] → synthesis → grade report 2
                                                ↓
                                      compile_gradebook
Up to concurrency: 3 papers run simultaneously. Within each paper, all three swrm agents run fully in parallel. The synthesis step fires once all three agents for that paper have finished. The factory’s writes path receives the collected list of per-paper grade reports.

​
Next steps

1000 Monkeys

Pure swrm fan-out: the same prompt sent to N agents simultaneously.

GitHub Issues Triage

Classic factory with a single agent per item.

Factory Reference

Full FactoryNode YAML reference including swrm mode.

Swrm Reference

Full swrm node documentation.