AI Agents
AI Agents
This is where AI stops being a conversation partner and starts being a colleague. An agent doesn’t just answer questions — it plans, acts, observes results, adjusts, and keeps going until the job is done.
You’re probably already using one. If you’ve used Claude Code, Cursor, or GitHub Copilot’s agent mode — that’s an AI agent. It reads your codebase, decides what to change, writes the code, runs the tests, and iterates on failures. Autonomously.
What Makes an Agent an Agent
A chatbot waits for you to type. An agent does things.
The difference is a loop:
1. OBSERVE → "Here's the task and current state"
2. THINK → "I should try this approach..."
3. ACT → Calls a tool (search, write file, run code, hit API)
4. OBSERVE → "That returned this result..."
5. THINK → "OK, now I need to..."
6. REPEAT → Until done (or stuck, or told to stop)The LLM provides the reasoning. Tools provide the actions. The loop provides the autonomy.
The Building Blocks
| Component | What it does | Why it matters |
|---|---|---|
| LLM brain | Reasons, plans, decides | The intelligence behind the decisions |
| Tools | Actions in the real world | Without tools, it’s just a chatbot |
| Memory | Tracks what’s happened | Short-term (context) + long-term (vector DB) |
| Planning | Breaks goals into steps | Complex tasks need decomposition |
| Guardrails | Safety boundaries | Permission systems, human approval gates |
How Tool Use Works
Modern LLMs are trained to output structured tool calls. When Claude wants to search the web, it doesn’t type a URL — it emits:
{"tool": "web_search", "query": "EU AI Act enforcement timeline 2025"}The system executes that tool, returns the result, and the model continues reasoning with new information. This is function calling, and it’s what makes agents possible.
Model Context Protocol (MCP)
Anthropic created MCP as an open standard — a universal plug for connecting AI to tools and data. Instead of building custom integrations for every tool, you build one MCP server and any compatible AI can use it.
Think of it like USB for AI tools. One standard, infinite possibilities.
Architectures
Single agent (ReAct) — One model reasons and acts in a loop. Simple, effective, the default.
Multi-agent — Specialised agents collaborate. One plans, another codes, another reviews. A coordinator orchestrates. More complex but handles harder problems.
Human-in-the-loop — Agent works autonomously but pauses at decision points for human approval. The responsible default for high-stakes work.
Agents You Can Use Today
| Agent | What it does |
|---|---|
| Claude Code | Reads codebases, plans changes, writes/tests code, terminal-native |
| Cursor / Windsurf | IDE-integrated coding agents |
| Devin | Autonomous software engineering |
| CrewAI | Multi-agent orchestration framework |
| LangGraph | Build custom agent workflows |
What Excites Me About This
Agents are the bridge between “AI that talks” and “AI that works.” They’re why AI coding tools feel magical — the model isn’t just suggesting text, it’s executing a plan.
But they’re also early. Current agents are brilliant for well-defined tasks (write this function, find this bug, research this topic) and unreliable for ambiguous ones. Knowing their limits is as important as knowing their strengths.
What I’m Still Learning
- How to design good tool interfaces (what makes a tool easy for an agent to use?)
- When to use a single agent vs multi-agent (where’s the complexity worth it?)
- How MCP changes what’s possible as the ecosystem grows
Go Deeper
- Prompt Engineering — How you instruct agent behaviour
- RAG & Retrieval — Agents often use RAG to ground their work in real data
- Tools & Frameworks — The practical tooling for building agents
- AI Alignment — Autonomy without alignment is dangerous
- How LLMs Work — Helps you understand what agents can and can’t reason about
Best Resources
- Anthropic “Building Effective Agents” — Best practical guide to agent architecture
- Andrew Ng’s “Agentic AI” talks — Strategic framing of where agents are heading
- LangChain / LangGraph docs — If you want to build one yourself
- Just use Claude Code — Honestly, using an agent daily teaches you more than reading about them