# Agents, harnesses, tools, and MCP

> Understand the parts behind a tool-using AI workflow.

Canonical: https://ainextia.com/learn-ai/agents-harnesses-mcp/
Publisher: https://ainextia.com/

Level: Intermediate
Track: How it works
Reviewed: 2026-11-05

## Part 1
An agent is a system that chooses actions toward a goal. A harness is the surrounding software that coordinates the model, tools, memory, permissions, evaluation, and runtime. MCP is a protocol for connecting models and agents to external tools and data.

## Part 2
The harness, not the model alone, determines which actions are actually possible and safe. Permission boundaries, approval points, and failure handling matter as much as the prompt.

## Part 3
Draw a flow diagram with user goal, model, tool calls, permissions, logs, review, and fallback.

## Knowledge check
Why doesn't giving an agent a new MCP server automatically make it reliable?

Reference: https://modelcontextprotocol.io/introduction
