Graph Config¶

Graph config controls workflow execution. Understanding key settings helps you build effective multi-agent workflows.

Configuration Overview¶

Graph config has two levels:

Level	What It Controls	Key Settings
Graph Level	Overall workflow	Name, description, output format
Node Level	Individual execution units	Agent/model, prompts, tools, connections

Graph Level Settings¶

Setting	Purpose	Notes
Name	Workflow identifier	Cannot contain `/` `\` `.`
Description	What this workflow does	Helps organize and search
End Template	Format final output	Combines results from multiple nodes

End Template combines node outputs using {{node_name}} syntax:

# Report
## Data: {{collector}}
## Analysis: {{analyzer}}

Node Settings¶

Each node has 6 configuration categories:

1. Identity¶

Setting	Required	Purpose
Name	Yes	Node identifier (no `/` `\` `.`)
Description	No	What this node does

2. Execution Source¶

Choose ONE approach:

Approach	When to Use	What You Get
Use Agent (`agent_name`)	Reuse configured agent	Inherits agent's model, prompts, tools
Use Model (`model_name`)	Custom one-off node	Full manual control over prompts and tools
Use Subgraph (`is_subgraph` + `subgraph_name`)	Call another workflow	Modular, reusable logic

Agent Override: When using an agent, node settings override agent defaults. For example, setting model_name in the node uses a different model while keeping the agent's prompts and tools.

3. Prompts¶

Setting	Purpose	Variable Support
System Prompt	Define node behavior and role	Yes
User Prompt	Specific task for this execution	Yes

Variable Substitution in prompts:

Syntax	Meaning	Example
`{{start}}`	User's original input	`Analyze: {{start}}`
`{{node_name}}`	Output from a node	`Summarize: {{researcher}}`
`{{node_name:3}}`	Last 3 outputs from node	`Review: {{checker:3}}`
`{{node_name:all}}`	All outputs from node	`Combine: {{processor:all}}`
`{{@prompt_name}}`	Reference registered prompt	`{{@data_analyst}}`

4. Tools¶

Setting	Purpose
System Tools	Built-in capabilities (`file_creator`, `memory_tools`, etc.)
MCP Servers	External services (web search, databases, APIs)

5. Execution Control¶

Setting	Default	Range	What It Does
Max Iterations	50	1-200	Limits tool calls per execution
Handoffs	-	-	How many times node can be re-selected (for loops)
Output Enabled	true	-	Show results to user vs silent processing

Max Iterations Guide:

Range	Use Case
1-10	Quick tasks, simple queries
10-30	Multi-step workflows
30-50	Complex analysis (default)
50-100	Deep research, code generation
100-200	Long-running automation

Handoffs enables loops. When set to 3, the node can be chosen up to 3 times before being disabled. See Handoffs for routing logic.

Output Enabled controls visibility:

Setting	Shows	Use For
`true`	Assistant response	User-facing results
`false`	Only tool results	Background processing

6. Connections¶

Setting	Purpose
Input Nodes	Where this node receives data from
Output Nodes	Where this node sends data to

Special Values:

Value	Meaning
`["start"]`	Receives user input
`["end"]`	Finishes workflow

Connection Examples:

Input	Output	Flow
`["start"]`	`["processor"]`	User → processor
`["collector"]`	`["analyzer"]`	collector → analyzer
`["validator"]`	`["end"]`	validator → finish
`["router"]`	`["option_a", "option_b"]`	router → multiple paths

Configuration Patterns¶

Quick Response¶

Max iterations: 10-20
Minimal tools
Single path (no branches)

Deep Analysis¶

Max iterations: 50-100
Multiple MCP servers
System tools for file/memory operations

Background Processing¶

Output enabled: false on processing nodes
Output enabled: true on final notification node

Modular Workflows¶

Use subgraphs for reusable logic
Each subgraph handles one responsibility
Connect subgraphs sequentially

Common Questions¶

Q: What's the difference between agent_name and model_name?

Setting	Uses	Best For
`agent_name`	Pre-configured agent	Reusable, consistent behavior
`model_name`	Direct model access	Custom, one-off tasks

Q: Can I override an agent's settings?

Yes. When using agent_name, any node setting overrides the agent default: - model_name → Different model - system_prompt → Different behavior - mcp_servers → Different tools - max_iterations → Different limit

Q: How do I reference previous node outputs?

Use {{node_name}} in prompts. The first node connecting to ["start"] can use {{start}} for user input.

Q: What happens when max_iterations is reached?

Execution stops gracefully. The node returns whatever was generated. No error occurs.

Q: Can subgraphs call each other?

Not in circles. If A calls B, then B cannot call A.

Q: How do nodes at the same level execute?

Sequentially, in the order they appear in the node list.