What is an LLM agent — the honest definition

Strip the marketing.

Agent = LLM that can:
1. Call tools (search, code, API)
2. Loop on its own outputs
3. Decide when to stop

That's it. Not a magic intelligence. Not 'AGI lite'. A program where the LLM is the planner, and external functions are the hands.

When this works:
- Multi-step research
- Long-form code generation with self-correction
- Browsing + reading + summarising

When it doesn't:
- Tasks needing precise determinism
- High-stakes single-shot decisions
- Tasks better served by a hard-coded pipeline

ReAct (Reasoning + Acting) is the dominant agent pattern.

For each step, the LLM produces:
1. Thought — what should I do next?
2. Action — which tool to call (or 'final answer')
3. Observation — the tool's result, fed back

Loop until 'final answer'. The chain of thoughts + actions + observations is the agent's trajectory.

Most frameworks (LangChain, LlamaIndex, AutoGen, CrewAI) implement ReAct under different names. The shape doesn't change — just the API.

No framework. Just chat completions + a tool dict + a while loop. Reading this once is worth 10 LangChain tutorials. Frameworks add memory, tracing, retries — but the core is what's below.

Agents are slow, expensive, and unpredictable.

Before building one, ask:
- Can this be a fixed pipeline (no LLM-driven branching)?
- Can a single LLM call with good prompting handle it?
- Can a tool-calling LLM (no loop) solve it?

Only if all three are 'no' does the loop earn its keep. Most teams should reach for agentic patterns at version 2, not version 1. Ship the dumb version first.

Day 71 done.

- Agent = LLM + tools + loop
- ReAct: thought / action / observation
- 30-line agent from scratch
- Reach for agents only when needed

Tomorrow (Day 72): tool design — the most important skill in agent engineering.