---
name: branching-agent-pattern
version: 1.0.0
description: Define and execute AI agent workflows using CSV-based declarative definitions
  with configurable branching logic
inputs:
- 'CSV workflow files defining agent graphs with columns: graph_name, node_name, agent_type,
  next_node, on_failure, prompt, input_fields, output_field'
- LLM provider configuration (OpenAI, Anthropic, Google) in agentmap_config.yaml
- Storage backend configuration in agentmap_config_storage.yaml
steps:
- Define workflow graph in CSV with nodes representing agent steps and their connections
  (next_node, on_failure)
- Configure BranchingAgent with customizable success/failure values and fallback fields
  in the context dictionary
- Initialize the agent runtime with ensure_initialized() and configure execution tracking
  and state adapter services
- Execute the workflow using agentmap run with appropriate inputs and monitor the
  execution trace
outputs:
- Executed workflow with results stored in the specified output_field
- Detailed execution trace showing success/failure decisions at each branching point
- Updated workflow state persisted in the configured storage backend
tags: []
metadata:
  source_repo: https://github.com/jwwelbor/AgentMap.git
  extracted_at: ''
  confidence: 0.95
---

# branching-agent-pattern

Define and execute AI agent workflows using CSV-based declarative definitions with configurable branching logic

## Setup

**Dependencies:**

```text
pip install langgraph>=1.0.0 langchain-core>=0.3.0 pyyaml>=6.0.0 fastapi>=0.111.0 uvicorn>=0.34.3
```

**Setup steps:**

1. Install AgentMap: pip install agentmap[all]
1. Configure llm providers in agentmap_config.yaml (OpenAI, Anthropic, Google models)
1. Create CSV workflow files with graph definitions
1. Initialize runtime with ensure_initialized()
1. Run workflow with agentmap run <csv_file> --pretty

## Key Files

- `agentmap_config.yaml - Main configuration with LLM and storage settings`
- `agentmap_config_storage.yaml - Storage backend configuration`
- `hello_world.csv - Sample workflow demonstrating basic agent chain`
- `examples/host_integration/custom_agents.py - Custom agent implementations with host service integration`

## Steps

1. Define workflow graph in CSV with nodes representing agent steps and their connections (next_node, on_failure)
2. Configure BranchingAgent with customizable success/failure values and fallback fields in the context dictionary
3. Initialize the agent runtime with ensure_initialized() and configure execution tracking and state adapter services
4. Execute the workflow using agentmap run with appropriate inputs and monitor the execution trace

## Implementation Details

```python
CSV format: graph_name,node_name,agent_type,next_node,on_failure,prompt,input_fields,output_field
```

```python
BranchingAgent context example: {'input_fields': ['success'], 'output_field': 'result', 'success_values': ['PASSED', 'COMPLETED']}
```

```python
Execution command: agentmap run hello_world.csv --pretty
```

## Inputs

- CSV workflow files defining agent graphs with columns: graph_name, node_name, agent_type, next_node, on_failure, prompt, input_fields, output_field
- LLM provider configuration (OpenAI, Anthropic, Google) in agentmap_config.yaml
- Storage backend configuration in agentmap_config_storage.yaml

## Outputs

- Executed workflow with results stored in the specified output_field
- Detailed execution trace showing success/failure decisions at each branching point
- Updated workflow state persisted in the configured storage backend

## Failure Modes

- Invalid CSV format causing parsing errors during workflow loading
- Missing or misconfigured LLM provider settings leading to execution failures
- Storage backend unavailable or misconfigured preventing workflow persistence
- Agent execution timeout due to long-running operations or infinite loops

## Source

Extracted from: [https://github.com/jwwelbor/AgentMap.git](https://github.com/jwwelbor/AgentMap.git)
Confidence: 0.95