Skip to main content

Agents & MCP

Agents​

An agent is an LLM in a loop: it picks a tool, observes the result, and repeats until the goal is done (or a limit is hit).

PatternUse when
Single callThe task fits one prompt - start here
Workflow / chainSteps are known in advance (extract → transform → write)
RouterClassify input, send to a specialized prompt or cheaper model
Agent loopSteps are unknown and depend on intermediate results
Orchestrator + subagentsBig tasks split into parallel, isolated subtasks

Agent rules of thumb:

  • Prefer the simplest pattern that works; agents cost more and fail in more ways.
  • Limit iterations, tokens, wall-clock time and spend.
  • Give the agent good tools (clear names, descriptions, error messages) - tool quality drives agent quality.
  • Human in the loop for irreversible actions.
  • Log every step (prompt, tool call, result) for debugging and evals.
  • Manage context: summarize or drop old tool output in long runs.

Model Context Protocol (MCP)​

An open protocol (JSON-RPC based) that standardizes how AI applications connect to tools and data. Write a server once; use it from any MCP-capable client (IDEs, chat apps, agent frameworks).

ConceptMeaning
Host / clientThe AI app (e.g. an IDE or chat app) that connects to servers
ServerYour process exposing capabilities
ToolsFunctions the model can call
ResourcesRead-only data the app can load into context (files, records)
PromptsReusable prompt templates exposed to users
Transportsstdio for local processes, Streamable HTTP for remote servers

Minimal MCP server (TypeScript)​

npm i @modelcontextprotocol/sdk zod
import { McpServer } from "@modelcontextprotocol/sdk/server/mcp.js";
import { StdioServerTransport } from "@modelcontextprotocol/sdk/server/stdio.js";
import { z } from "zod";

const server = new McpServer({ name: "notes", version: "1.0.0" });

server.registerTool(
  "search_notes",
  {
    title: "Search notes",
    description: "Full-text search over my notes. Returns matching titles.",
    inputSchema: { query: z.string() },
  },
  async ({ query }) => ({
    content: [{ type: "text", text: JSON.stringify(await searchNotes(query)) }],
  }),
);

await server.connect(new StdioServerTransport());
  • Python: the official mcp package on PyPI (check its README for the current server API - it is on a 2.x major).
  • Debug servers with the MCP Inspector: npx @modelcontextprotocol/inspector.

MCP security​

  • A server runs with your permissions; install only trusted servers.
  • Tool descriptions and tool outputs go into the model's context, so a malicious server can inject instructions.
  • Use least-privilege credentials per server and require confirmation for writes.
  • For remote servers, use the spec's OAuth-based authorization rather than shared static keys.