What’s your first association with the word “agent”? 007? Smith from The Matrix? Mine’s J and K from Men in Black! So every time “AI agent” comes up on a call, I mentally reply: name, and how many aliens this week.
But on to AI agents. What are they, and what’s the difference between an agent, MCP, and tool use?
LLM: generates text based on what it already knows Tool use: lets the model call external functions and act, not just write MCP: a standard for how tools connect to models. Think USB-C for AI
Let’s dive a little into MiB world. Recruit J gets an assignment: find out why alien activity is up in sector Z.
→ J as a pure LLM: writes a report from what he remembers at the academy. Done.
→ J with tool use: googles sector Z, finds an article about a honey farm nearby, writes a report. Done.
→ Now let’s promote J to agent with the same task assigned.
Agent J plans the work himself. Runs a sector search. Decides from the data how to get there. Sees the aliens are armed. Adjusts the plan. Comes back, picks up subagents: K for translation, I for backup. Confirms the honey farm is in fact the cause. Returns with two options: shut down the farm, or send the aliens back to their home planet.
An agent, in the engineering sense, is three things:
- Loop: act, observe, decide, act again
- Tools: what the agent uses to affect its environment
- Autonomy: it picks each next step on its own, until the goal is reached
Claude Code is an agent. Give it a task: it runs in a loop, calls tools, spawns subagents when it needs to, and keeps going until the job is closed.
Subagents are K and I. They don’t need J’s full context. Just their piece of the task and a result to bring back.
So next time someone says “I built an agent” - ask how many it’s caught. 😉
An agent doesn’t really “know” it’s an agent. The model is the same. The agency lives in the loop around it.
