Creating Custom Agents

Overview

While MINDBLIX provides 100+ pre-built agent templates, you may need custom agents for specialized tasks. This guide walks you through creating one from scratch.

Agent Definition

Every custom agent requires three components:

System Prompt

Defines the agent’s personality, expertise, and behavior rules.

Tool Configuration

Specifies which external tools the agent can use.

Output Schema

Defines the structure of the agent’s output.

Step-by-Step Guide

Define the Agent's Purpose

Start by clearly defining what your agent should do:

AGENT_CONFIG = {
    "name": "Patent Analyzer",
    "description": "Analyzes patent documents for novelty and prior art",
    "domain": "legal",
    "version": "1.0.0"
}

Write the System Prompt

The system prompt shapes the agent’s behavior:

SYSTEM_PROMPT = """
You are a Patent Analysis Agent specialized in:
- Analyzing patent claims for novelty
- Identifying prior art references
- Evaluating patent strength and validity
- Generating patent landscape reports

Rules:
- Always cite specific patent numbers when referencing prior art
- Provide confidence scores for all assessments
- Output in structured JSON format
- Flag potential conflicts of interest
"""

Configure Tools

Define which tools your agent can access:

TOOLS = [
    {
        "name": "patent_search",
        "description": "Search patent databases (USPTO, EPO, WIPO)",
        "parameters": {
            "query": "string",
            "jurisdiction": "string",
            "date_range": "object"
        }
    },
    {
        "name": "document_reader",
        "description": "Extract text from PDF patent documents",
        "parameters": {
            "file_path": "string"
        }
    }
]

Define Output Schema

Specify the expected output structure:

{
  "type": "object",
  "properties": {
    "novelty_score": { "type": "number", "min": 0, "max": 1 },
    "prior_art": {
      "type": "array",
      "items": {
        "patent_number": "string",
        "relevance": "number",
        "summary": "string"
      }
    },
    "recommendation": { "type": "string", "enum": ["file", "revise", "abandon"] },
    "confidence": { "type": "number" }
  }
}

Deploy

Deploy your agent via the API:

curl -X POST http://localhost:8000/api/v1/agents \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d @patent-analyzer-config.json

Best Practices

Keep prompts focused

Each agent should do one thing well. If your prompt is trying to handle multiple domains, consider splitting into multiple agents.

Define clear output schemas

Structured output makes it easier to chain agents together in workflows and parse results programmatically.

Test with edge cases

Test your agent with ambiguous inputs, missing data, and adversarial prompts to ensure robustness.

Version your agents

Always version your agent configurations so you can roll back if a new version doesn’t perform as expected.

Getting Started

Core Concepts

Guides

Agent Catalog

Overview

Agent Definition

System Prompt

Tool Configuration

Output Schema

Step-by-Step Guide

Best Practices

​Overview

​Agent Definition

System Prompt

Tool Configuration

Output Schema

​Step-by-Step Guide

​Best Practices

Overview

Agent Definition

Step-by-Step Guide

Best Practices