Agent Pattern Configuration

Agent patterns define the execution strategies available to the agent orchestrator. The orchestrator’s meta-router uses these definitions to select the most appropriate reasoning pattern for each incoming question.

Overview

TrustGraph ships with three built-in agent patterns:

  • react - Interleaved reasoning and action (ReAct). Iteratively thinks, selects a tool, observes the result, and repeats until it can answer. Good for open-ended, exploratory tasks.
  • plan-then-execute - Generates a multi-step plan first, then executes each step sequentially and synthesises the results. Suited to structured, predictable tasks.
  • supervisor - Decomposes the question into independent sub-agent goals, fans out to parallel sub-agents, and aggregates findings into a synthesis. Best for broad, multi-dimensional analysis.

Patterns are stored in the configuration service with the type agent-pattern and are managed through the standard configuration CLI commands.

JSON Structure

Each pattern is a JSON object with the following fields:

{
  "name": "react",
  "description": "ReACT \u2014 Reasoning + Acting",
  "when_to_use": "Adaptive, good for open-ended tasks"
}
Property Required Description
name Yes Pattern identifier, must match the config key
description Yes Human-readable description presented to the LLM meta-router
when_to_use No Additional guidance for the meta-router on when to select this pattern

Default Behaviour

If no agent-pattern config items are defined, the orchestrator falls back to a single pattern:

{
  "name": "react",
  "description": "Interleaved reasoning and action"
}

Managing Patterns

Adding a Pattern 

tg-put-config-item --type agent-pattern --key supervisor --value '{
  "name": "supervisor",
  "description": "Fan-out to parallel sub-agents with synthesis",
  "when_to_use": "Broad questions requiring multi-dimensional analysis"
}'

Listing Patterns 

tg-list-config-items --type agent-pattern

Retrieving a Pattern 

tg-get-config-item --type agent-pattern --key react

Deleting a Pattern 

tg-delete-config-item --type agent-pattern --key supervisor

How Patterns Are Selected

The agent orchestrator’s meta-router selects a pattern in two stages:

  1. Task type identification - The LLM classifies the incoming question into a known task type (see Agent Task Types)
  2. Pattern selection - The LLM selects from the patterns listed in the task type’s valid_patterns field, using each pattern’s description and when_to_use to inform the decision

The selected pattern can also be specified explicitly using the -p flag on the CLI:

tg-invoke-agent -q "Analyse the risks" -p supervisor

See Also