Agents & Tool Use

An agent is an LLM that can call functions. That single capability — calling code — turns a chatbot into a programmable system.

The loop

while not done:
  1. LLM looks at conversation + available tools
  2. LLM decides: respond, OR call a tool
  3. If tool call: run it, append result to conversation
  4. Else: return final answer

Why it's powerful

• Real data. The LLM can call get_weather() instead of guessing.
• Real actions. It can send an email, book a meeting, file a PR.
• Long horizons. It can run for hours, decomposing a goal into many tool calls.

Why it's dangerous

• Loops. Agents get stuck repeating themselves. Always set a step budget.
• Hallucinated tool calls. Validate every argument before executing.
• Cost explosion. Each step is another LLM call. A 50-step agent on a paid model isn't free.

Patterns

• ReAct (Reason + Act) — the OG pattern; the LLM verbalises its plan, then acts.
• Plan-and-Execute — generate the full plan up front, then execute deterministically.
• Multi-agent — multiple specialised agents (planner, coder, critic) that hand work off.

Most production agents are simpler than the hype. They usually have 3-5 tools and a tight loop budget.