AI Agents & MCP

A chat model tells you what to do. An agent does it. MCP is how it does it.

Agent Components

100s

MCP Servers

Major Platforms

∞

Real-World Tasks

The agent loop

while not done:

      thought = model.think(history + tools)

      // pick a tool, or finish

      action = thought.next_tool_call

      observation = tools.execute(action)

      history.append(thought, action, observation)

The 5 components

REASONING

1. Model

Opus 4.7, GPT-5, Gemini Ultra. Smart enough to plan + self-correct.

2. Tools

Functions the agent can call. Each has a name, description, parameter schema.

3. Loop / harness

The orchestration code asking model what's next, executing, returning.

4. Memory

Working memory (this conversation) + optional long-term (RAG, vectors).

5. Safety / oversight

Confirmation gates, allow/deny lists, sandboxing, step budgets. Human-in-the-loop on destructive actions.

What is MCP?

Model Context Protocol — the open standard for connecting AI agents to external tools. USB for AI. Anthropic created it, every major vendor adopted it.

Top Failure Modes

Goal drift · Loop divergence · Hallucinated tool calls · Costly tangents · Prompt injection from external data

Pick by use case

High-value agent tasks

Sustained coding
→ Claude Code
Browser automation
→ ChatGPT Operator
Google Workspace tasks
→ Gemini
Real-time X / news
→ SuperGrok

Multi-file refactors
Across your codebase
App Store metadata
Build status, release notes
Cross-tool workflows
GitHub → Linear → Slack
Continuous monitors
Morning email summary