Agent Task Type Configuration

Agent task types define categories of work that the agent orchestrator can recognise and handle with specialised framing. When a question arrives, the meta-router classifies it into a task type, applies domain-specific framing to the LLM context, and constrains which agent patterns are available.

Overview

Task types allow you to:

  • Classify questions into domain categories (e.g. research, risk assessment, summarisation)
  • Inject framing prompts that steer the LLM’s approach for that category
  • Constrain pattern selection so that only suitable execution patterns are considered

Task types are stored in the configuration service with the type agent-task-type and are managed through the standard configuration CLI commands.

JSON Structure

Each task type is a JSON object with the following fields:

{
  "name": "risk-assessment",
  "description": "Due Diligence / Risk Assessment",
  "framing": "Analyse across financial, reputational, legal and operational dimensions using structured analytic techniques.",
  "valid_patterns": ["supervisor", "plan-then-execute", "react"],
  "when_to_use": "Multi-dimensional analysis requiring structured assessment"
}
Property Required Description
name Yes Task type identifier, must match the config key
description Yes Human-readable description presented to the LLM meta-router for classification
framing No Domain-specific framing prompt injected into the LLM context when this task type is selected
valid_patterns No List of agent pattern keys that are allowed for this task type. If omitted, all patterns are available
when_to_use No Additional guidance for the meta-router on when to classify a question as this task type

Default Behaviour

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

{
  "name": "general",
  "description": "General queries",
  "valid_patterns": ["react"],
  "framing": ""
}

Managing Task Types

Adding a Task Type 

tg-put-config-item --type agent-task-type --key research --value '{
  "name": "research",
  "description": "In-depth research and investigation",
  "framing": "Conduct thorough research, cross-reference multiple sources, and provide well-evidenced conclusions.",
  "valid_patterns": ["plan-then-execute", "react"],
  "when_to_use": "Questions requiring detailed investigation across multiple topics"
}'

Listing Task Types 

tg-list-config-items --type agent-task-type

Retrieving a Task Type 

tg-get-config-item --type agent-task-type --key research

Deleting a Task Type 

tg-delete-config-item --type agent-task-type --key research

Example Configuration

A typical deployment might define several task types:

# General-purpose queries - use ReAct for flexibility
tg-put-config-item --type agent-task-type --key general --value '{
  "name": "general",
  "description": "General queries and simple questions",
  "framing": "",
  "valid_patterns": ["react"]
}'

# Research tasks - plan first, then execute
tg-put-config-item --type agent-task-type --key research --value '{
  "name": "research",
  "description": "In-depth research and investigation",
  "framing": "Conduct thorough research, cross-reference multiple sources, and provide well-evidenced conclusions.",
  "valid_patterns": ["plan-then-execute", "react"]
}'

# Risk assessment - fan out to parallel sub-agents
tg-put-config-item --type agent-task-type --key risk-assessment --value '{
  "name": "risk-assessment",
  "description": "Due Diligence / Risk Assessment",
  "framing": "Analyse across financial, reputational, legal and operational dimensions using structured analytic techniques.",
  "valid_patterns": ["supervisor", "plan-then-execute"],
  "when_to_use": "Multi-dimensional analysis requiring structured assessment"
}'

How Task Types Are Used

When the agent orchestrator receives a question:

  1. The meta-router presents all configured task types (with their description and when_to_use fields) to the LLM
  2. The LLM classifies the question into the most appropriate task type
  3. The task type’s framing prompt is injected into the agent’s LLM context
  4. Pattern selection is constrained to the task type’s valid_patterns list
  5. The meta-router selects the best pattern from the constrained set

If classification fails, the orchestrator falls back to the general task type with the react pattern.

See Also