Workflows

August 12, 2025 · View on GitHub

Workflows in Nerve allow you to chain multiple agents in a simple, linear sequence. Each agent in a workflow executes in order, passing shared state (variables) to the next. This is ideal for tasks that can be split into clear, sequential steps.

📌 For more advanced orchestration (parallel execution, sub-agents, branching logic), it's recommended to use MCP and expose agents as tools to a primary orchestrator.

📋 Example: Recipe Workflow

In this example, a workflow chains four agents to:

  1. Generate a list of ingredients
  2. Describe preparation steps
  3. Estimate prep time
  4. Rewrite the result in an engaging way

You can run it with:

nerve run examples/recipe-workflow --food pizza

YAML Definition

name: "Write a Recipe"
description: "A workflow for writing a recipe."

actors:
  create_list_of_ingredients:
    generator: anthropic://claude

  describe_preparation_steps:
    generator: openai://gpt-4o

  estimate_time:
    generator: openai://gpt-4o-mini

  rewrite_nicely:
    generator: openai://gpt-4o

Each key in actors defines an agent in the pipeline, executed in sequence.

🧠 Agent Files

Each agent is a standalone YAML with a task and tools. These are simplified below:

create_list_of_ingredients.yml

agent: You are a talented chef.
task: prepare a list of ingredients for {{ food }}
using:
  - reasoning

tools:
  - name: create_list_of_ingredients
    description: Provide ingredients, one per line.
    complete_task: true
    arguments:
      - name: ingredients
        description: Ingredients list.
        example: >
          - 1 cup flour
          - 2 eggs

describe_preparation_steps.yml

agent: You are a creative chef.
task: describe the preparation steps for {{ food }} using {{ ingredients }}
using:
  - reasoning

tools:
  - name: describe_preparation_steps
    description: Preparation steps.
    complete_task: true
    arguments:
      - name: steps
        example: >
          - Mix ingredients
          - Bake for 20 minutes

estimate_time.yml

agent: Estimate total prep time using the given ingredients and steps.
task: Estimate time to prepare {{ food }} using:
  {{ ingredients }}
  {{ steps }}
using:
  - reasoning

tools:
  - name: estimate_time
    description: Estimated preparation time
    complete_task: true
    arguments:
      - name: preparation_time
        example: 25 minutes

rewrite_nicely.yml

agent: You are a food blogger.
task: Rewrite the recipe for {{ food }} in an engaging way.

Includes:
  - {{ preparation_time }}
  - {{ ingredients }}
  - {{ steps }}

tools:
  - name: rewrite
    description: Final blog-ready recipe
    complete_task: true
    arguments:
      - name: recipe
        example: >
          # **Epic Lazy Day Pancakes**
          [ ... ]

🔄 State and Variables

Variable Passing

Variables are shared across all agents in a workflow. Each agent can:

  • Read variables set by previous agents
  • Set new variables for subsequent agents
  • Override existing variables

Default Values

Each agent can define defaults that apply when variables aren't set:

defaults:
  temperature: 0.7
  max_tokens: 1000

Variable Interpolation

All text fields support Jinja2 templating:

task: Process {{ food }} with {{ ingredients }}

📎 Notes

  • Agents receive inputs from previous agents via templating variables (e.g., {{ ingredients }})
  • Each tool must call complete_task: true to advance the workflow
  • All agents can define their own generators, tools, and prompts independently