Blind Code Review

A coder writes a function from a plain-English spec. A reviewer critiques the code without seeing the original spec — only the implementation. The coder then revises based on the feedback. The reviewer’s blindness makes the critique reflect only what the code communicates on its own.

What it demonstrates

Multi-turn refinement within a linear sequential pipeline
Agent specialization: one writes, one reviews, one revises
Context accumulation: working.code_v1 and working.review are both visible to the final node
How to model a real-world workflow (write → review → revise) in SirenSpec

Run it

sirenspec run docs/cookbook/blind-code-review/workflow.yaml
# Try a different problem:
sirenspec run docs/cookbook/blind-code-review/workflow.yaml \
  --input "Write a Python function that checks whether a string is a valid palindrome."

Check working.code_v1, working.review, and output.code_final in the trace to follow the refinement.

Workflow

docs/cookbook/blind-code-review/workflow.yaml

version: "0.1"

agents:
  coder:
    model: "openai:gpt-4o-mini"
    system: |
      You are a software engineer. Write a clean Python implementation of the
      described function. Include a docstring. Return only the code.

  reviewer:
    model: "anthropic:claude-haiku-4-5-20251001"
    system: |
      You are a senior engineer doing a code review. You have NOT seen the original
      spec — evaluate the code solely on what it does. Cover: correctness, edge cases,
      readability, and one concrete improvement suggestion. Under 150 words.

  coder_revised:
    model: "openai:gpt-4o-mini"
    system: |
      You submitted a function for review and received feedback. Incorporate the
      reviewer's suggestions and return a revised Python implementation. Code only.

nodes:
  write_code:
    agent: coder
    writes: working.code_v1
  review_code:
    agent: reviewer
    writes: working.review
  revise_code:
    agent: coder_revised
    writes: output.code_final

edges:
  - from: write_code
    to: review_code
  - from: review_code
    to: revise_code

guardrails:
  - injection
  - length

Basics

Multi-Agent Patterns

Swrm & Fan-Out

Factory & Iteration

Guardrails

Human in the Loop

Budget & Cost Control

Stress Tests

Tool Nodes

What it demonstrates

Run it

Workflow

Graph

Next steps

Graphic Design Firm

PR Summarizer

​What it demonstrates

​Run it

​Workflow

​Graph

​Next steps

Graphic Design Firm

PR Summarizer

What it demonstrates

Run it

Workflow

Graph

Next steps