Agent Harness

AI agent harness for small language models (SLMs) with a lightweight tool calling protocol.

Overview

A minimal, token-efficient agent loop designed for SLMs (3B-8B parameters). Built on Ollama for local inference.

Architecture

┌─────────────────────────────────────┐
│           Tool Registry             │
│  (definitions + execution engine)  │
└──────────────┬──────────────────────┘
               │
┌──────────────▼──────────────────────┐
│         Prompt Composer             │
│  (system + tools + history + query) │
└──────────────┬──────────────────────┘
               │
┌──────────────▼──────────────────────┐
│        Model Backend (Ollama)       │
└──────────────┬──────────────────────┘
               │
┌──────────────▼──────────────────────┐
│        Output Parser                │
│   (extract tool calls / responses)  │
└─────────────────────────────────────┘

Quick Start

npm install
npm run dev

Configuration

Edit config.json to set:

• model: Ollama model to use
• tools: Enabled tools
• maxIterations: Max tool calls per turn
• ollamaUrl: Ollama endpoint

Tool Protocol

Tools are defined as JSON schemas:

{
  "name": "filesystem_read",
  "description": "Read contents of a file",
  "parameters": {
    "type": "object",
    "properties": {
      "path": { "type": "string", "description": "File path to read" }
    },
    "required": ["path"]
  }
}

Model outputs tool calls in this format:

<tool_call>
{"name": "filesystem_read", "arguments": {"path": "/etc/hosts"}}
</tool_call>

Available Tools

• filesystem_read - Read files
• filesystem_write - Write files
• bash - Execute shell commands
• web_search - Search the web (Brave API)
• memory_search - Query persistent memory

License

MIT