Agent Templates

Agent Templates are the core configuration mechanism in AgentDock, allowing you to define an agent's identity, capabilities, and behavior in a declarative JSON format.

Overview

Each agent in the /agents directory has a template.json file. This file defines:

Basic identity (ID, name, description)
LLM provider and model selection, including model parameters
System prompt and personality traits
Available tools
Orchestration rules (steps, conditions, sequences)

This templating system makes it easy to create, share, and modify agents without writing extensive code.

See Contributing Community Agents for information on how to add your own agent templates to the public repository.

Template Structure (`template.json`)

The template.json file follows this general structure:

{
  "version": "1.0", // Optional: Version of the template format
  "agentId": "unique-agent-id", // Required: Unique identifier
  "name": "Display Name", // Required: Name shown in UI
  "description": "Brief description of the agent.", // Required: Description for UI
  "tags": ["Example", "Research"], // Optional: Tags for categorization
  "priority": 10, // Optional: Lower numbers appear higher in lists
  "personality": [ // Required: System prompt broken into lines/paragraphs
    "Personality trait 1",
    "Personality trait 2"
  ],
  "nodes": [ // Required: List of node types used by the agent
    "llm.openai", // Example: Specify the LLM node provider
    "search"      // Example: Include necessary tool nodes
  ],
  "nodeConfigurations": { // Required: Configuration for specific nodes
    "llm.openai": { // Key matches the node type from the "nodes" list
      "model": "YOUR_CHOSEN_MODEL", // Required: Specify the model ID
      "temperature": 0.7, // Optional: Controls randomness (0=deterministic, >0=more random). Range varies by provider.
      "maxTokens": 4096, // Optional: Max tokens for the response.
      "topP": 0.9, // Optional: Nucleus sampling (0-1). Consider only top P% probability mass. Use temperature OR topP.
      "topK": 50, // Optional: Consider only the top K most likely tokens.
      "frequencyPenalty": 0.2, // Optional: Penalizes frequently used tokens (0=no penalty). Range varies.
      "presencePenalty": 0.1, // Optional: Penalizes tokens already present in prompt/response (0=no penalty). Range varies.
      "stopSequences": ["\nUser:"], // Optional: Sequences that stop generation.
      "seed": 12345, // Optional: Integer for deterministic results (if supported).
      "useCustomApiKey": false // Optional: If true, requires user to provide API key in settings.
    },
    "search": { // Example: Configuration for a tool node (if needed)
      "maxResults": 5 
    }
  },
  "chatSettings": { // Required: Settings for the chat interface
    "historyPolicy": "lastN", // Optional: 'none', 'lastN', 'all' (default: 'lastN')
    "historyLength": 20, // Optional: Number of messages if policy is 'lastN' (default: 50)
    "initialMessages": [ // Optional: Messages shown when chat starts
      "Hello! How can I help?"
    ],
    "chatPrompts": [ // Optional: Suggested prompts shown in UI
      "What can you do?"
    ]
  },
  "options": { // Optional: Additional agent-level options
    "maxSteps": 10 // Example: Max tool execution steps per turn
  }
}

Key Configuration Fields

agentId, name, description: Basic identification.
personality: Defines the system prompt and core behavior. Crucial for guiding the LLM.
nodes: Lists all capabilities (LLM provider node, tool nodes) the agent requires.
nodeConfigurations: Allows setting specific parameters for each node listed in nodes.
- For LLM nodes (e.g., llm.openai), you must specify the model.
- You can optionally override default LLM behavior by setting parameters like temperature, maxTokens, topP, topK, frequencyPenalty, presencePenalty, stopSequences, and seed. The exact behavior and valid ranges for these settings can vary between different LLM providers.
chatSettings: Controls the user interface behavior, initial state, and prompt suggestions.

For detailed explanations of the common LLM settings (temperature, topP, maxTokens, etc.) and their effects, refer to the Vercel AI SDK Settings Documentation.

Featured Agents

Agent	Description	GitHub
Cognitive Reasoner	Tackles complex problems using a suite of cognitive enhancement tools and structured reasoning. Features different operational modes including Research, Problem-Solving, Evaluation, Comparison, Ideation, and Debate. Uses the think tool for step-by-step reasoning.	View Code
Dr. House	Medical diagnostician inspired by the TV character, specializing in advanced medical diagnostics and rare disease identification. Leverages comprehensive medical knowledge and medical databases.	View Code
Science Translator	Makes complex scientific papers accessible by finding and translating them into simple language without sacrificing accuracy. Utilizes PubMed access and multi-database scientific research capabilities.	View Code
Calorie Vision	Analyzes food images to provide precise calorie and nutrient breakdowns using visual recognition technology. Integrates with visual analysis tools to process and evaluate food content from photos.	View Code
Harvey Specter	Legal strategist and negotiator inspired by the Suits character, specializing in contract review, case strategy, and negotiation tactics. Accesses legal databases to provide accurate and actionable legal insights.	View Code
Orchestrated Agent	Demonstrates advanced agent orchestration by combining multiple specialized agents and tools in a cohesive workflow with dynamic branching. Shows how different agents can be combined and orchestrated.	View Code
Agent Planner	Specialized agent for designing and implementing AI agents using the AgentDock framework and RFA system. Provides agent ideation, architecture design, implementation guidance, and RFA system integration.	View Code
Tenant Rights Advisor	Guides renters through housing issues like repairs, evictions, and deposit disputes based on general housing regulations.	View Code
Consumer Rights Defender	Helps consumers navigate issues with refunds, warranties, defective products, and unfair billing practices.	View Code
Small Claims Court Guide	Assists with small claims court navigation including filing paperwork, preparing evidence, and collecting judgments.	View Code

Agent File Structure

For detailed implementation examples, clone the AgentDock repository and explore the agents/ directory.

agents/
└── agent-name/
    ├── template.json     # Core configuration (required)
    ├── README.md         # Documentation (recommended)
    └── assets/           # Optional assets (e.g., avatar.png)

Usage

These templates can be:

Tested directly: Try them out via the chat interface by clicking the agent links above (if running the client).
Examined for patterns: Study their implementations to learn configuration techniques.
Used as starting points: Copy and modify them to create your own specialized agents.