Agentform Examples

The examples/ directory contains ready-to-use configurations demonstrating various Agentform features. Each example includes detailed documentation explaining the concepts it covers.

Examples Overview

Simple Agent

Directory: simple-agent/

The simplest possible Agentform configuration. A single agent that answers questions using OpenAI.

What You’ll Learn

• Basic Agentform file structure with numbered prefixes
• Defining variables, providers, and models
• Creating an agent with instructions
• Building a simple workflow

Features Used

• Variables with sensitive = true
• LLM provider configuration
• Policy budgets (cost and timeout)
• Fallback models
• Single-step LLM workflow

Run It

1
2
cd examples/simple-agent
agentform run ask --var openai_api_key=$OPENAI_API_KEY --input '{"question": "What is Agentform™?"}'

Multi-Agent

Directory: multi-agent/

Multiple specialized agents working together with intelligent task routing.

What You’ll Learn

• Using multiple LLM providers (OpenAI + Anthropic)
• Creating specialized agents for different tasks
• Conditional workflow routing based on classification
• Differentiated policies for fast vs. thorough processing

Features Used

• Multiple providers (OpenAI and Anthropic)
• Multiple models across providers
• Different policies (fast vs thorough)
• Classifier agent for task routing
• Conditional steps (type = "condition")
• State passing between steps

Run It

1
2
3
4
5
6
7
8
9
10
11
12
13
cd examples/multi-agent

# Simple task → quick_responder (GPT-4o)
agentform run smart_respond \
  --var openai_api_key=$OPENAI_API_KEY \
  --var anthropic_api_key=$ANTHROPIC_API_KEY \
  --input '{"task": "What is 2+2?"}'

# Complex task → deep_analyst (Claude)
agentform run smart_respond \
  --var openai_api_key=$OPENAI_API_KEY \
  --var anthropic_api_key=$ANTHROPIC_API_KEY \
  --input '{"task": "Analyze the impact of AI on employment"}'

Filesystem Agent

Directory: filesystem-agent/

An agent that interacts with the filesystem via MCP (Model Context Protocol) server.

What You’ll Learn

• Integrating MCP servers for external capabilities
• Defining capabilities from server methods
• Using call steps to invoke capabilities
• Human approval for write operations
• Capability call budgets

Features Used

• MCP server configuration (server block)
• Capability definitions with side effects
• requires_approval = true for sensitive operations
• type = "call" workflow steps
• max_capability_calls budget
• Multiple workflows in one project

Run It

1
2
3
4
5
6
7
8
9
10
11
cd examples/filesystem-agent

# Read and summarize a file
agentform run read_and_summarize \
  --var openai_api_key=$OPENAI_API_KEY \
  --input '{"file_path": "article.txt", "task": "Summarize this"}'

# List directory contents
agentform run list_and_read \
  --var openai_api_key=$OPENAI_API_KEY \
  --input '{"directory_path": ".", "task": "What files are here?"}'

PR Reviewer

Directory: pr-reviewer/

Automated GitHub pull request reviewer with human-in-the-loop approval.

What You’ll Learn

• GitHub MCP integration with authentication
• Multi-step data gathering workflows
• Human approval gates before write operations
• Building context from multiple API calls
• Safe patterns for automated actions

Features Used

• MCP server with auth block for tokens
• Capabilities: read (get_pr, list_files) and write (create_review)
• requires_approval = true on write capabilities
• type = "human_approval" workflow step
• Sequential capability calls to build context
• Passing state between steps

Run It

1
2
3
4
5
6
cd examples/pr-reviewer

agentform run review_pr \
  --var openai_api_key=$OPENAI_API_KEY \
  --var github_personal_access_token=$GITHUB_TOKEN \
  --input '{"owner": "your-org", "repo": "your-repo", "pr_number": 123}'

The workflow will pause for your approval before submitting the review to GitHub.

File Structure Convention

All examples follow the same file naming convention:

1
2
3
4
5
6
7
8
9
10
11
example/
├── 00-project.agentform      # Project metadata (agentform block)
├── 01-variables.agentform    # Variable definitions
├── 02-providers.agentform    # Provider and model definitions
├── 03-servers.agentform      # MCP server configuration (if needed)
├── 04-capabilities.agentform # Capability definitions (if needed)
├── 05-policies.agentform     # Policy definitions
├── 06-agents.agentform       # Agent definitions
├── 07-workflows.agentform    # Workflow definitions
├── input.yaml          # Sample input
└── README.md           # Detailed documentation

Numbered prefixes ensure files are processed in the correct order. References work across files—for example, agents can reference models defined in the providers file.

Next Steps

After exploring these examples:

1. Create your own agent: Start with simple-agent as a template
2. Add capabilities: Use filesystem-agent as a guide for MCP integration
3. Build complex workflows: Use multi-agent patterns for routing