LLM Agents: Tools in a Loop

📖 Quick Start

• Who it’s for: Anyone curious about AI coding agents
• Time to complete: 5 minutes
• Prerequisites: None
• Expected outcome: Clear mental model of how agents work
• Next step: Try Amp in 15 minutes

Inspired by Simon Willison’s definition—“LLMs calling tools in a loop to achieve a goal”—this post keeps things simple and practical. If you remember one thing about agents, make it this: an agent is a loop that thinks, acts with tools, observes, and repeats until done.

What is an Agent?

• An agent uses an LLM to decide what to do next
• It can call tools (functions/APIs/scripts) to act in the world
• It runs in a loop: think → act → observe → decide → repeat, guided by a goal

The Core Loop

goal → think → choose_tool → call_tool → observe_result
    → decide: done? otherwise refine plan and loop

flowchart TD
    Start([Goal]) --> Think[Think: Analyze situation]
    Think --> Choose[Choose Tool]
    Choose --> Call[Call Tool]
    Call --> Observe[Observe Result]
    Observe --> Decide{Goal Reached?}
    Decide -->|No| Think
    Decide -->|Yes| End([Done])

Tools an Agent Can Call

• Files: read/write files, list directories
• Code: run a snippet (e.g., Python/TypeScript) or a test
• HTTP/APIs: fetch JSON, call app backends, web search
• Data: query a database or vector index
• System: shell commands or task runners (with guardrails)

A Quick Example

Goal: “Turn expenses.csv into a bar chart of monthly totals.”

1. Think: “I need to read the file, group by month, and plot.”
2. Act: Call file.read('expenses.csv') → observe rows
3. Think: “I’ll write and run a small Python snippet.”
4. Act: Call code.run(python, script) → observe a PNG path
5. Decide: Goal reached → return “Chart saved at charts/monthly.png”

🔨 Try It Now (2 minutes)

Task: Watch an agent think, act, and verify in your own codebase

Prompt:

Read package.json, list all dependencies, and create a summary
markdown file called deps-summary.md with total count.

Verification:

• Check that deps-summary.md was created
• Verify the dependency count matches package.json

Expected outcome: You’ll see the agent read the file, count dependencies, create a new file, and confirm it worked—the core loop in action.

When Does the Loop Stop?

• Goal reached: the result satisfies the instruction
• Safety limits: max steps/tokens/time reached
• Stuck: repeated failures or no progress detected
• User stop: human cancels or asks for partial results

Why This Matters

• Moves beyond chat: the LLM can actually do things via tools
• Robustness via iteration: adjust after each observation
• Composability: clear, testable tools keep agents maintainable
• Safety by design: constrain tools, log actions, cap loops

Quick FAQ

Q: How is an agent different from ChatGPT or Copilot? A: ChatGPT gives you text responses. Copilot suggests code. Agents run tools in a loop—they read files, run tests, see results, and iterate until the goal is met.

Q: What if the agent makes a mistake? A: Review diffs, stage good changes, discard bad ones. Git is your safety net.

Q: Can agents write entire apps? A: They excel at focused tasks with clear goals. Think “fix this bug and verify” not “build me Instagram.”

What’s Next

Now that you understand the core loop, it’s time to experience it yourself:

Next: Get Your First Win in 15 Minutes — Watch the agent loop in action on a real task in your codebase.

Alternative start: Coding with Agents in 2025 — Complete overview and mindset before diving in.

Practice Path:

1. You are here: What is an Agent
2. First Win in 15 Minutes
3. Workflows That Stick
4. Power Patterns
5. Planning Workflow

---

Reference: Simon Willison’s post on agents provides the foundational definition this intro builds on.