Tools vs Agents: Choosing the Right Abstraction

The Three Levels of AI Interaction

Prompt — a single LLM call. Fast, cheap, stateless. Right for: classification, extraction, generation, summarization.

Tool-augmented prompt — an LLM call that can invoke external functions. Right for: tasks requiring real-time data, computation, or external API calls.

Agent — a persistent, goal-directed system with memory, planning, and multiple tool calls. Right for: multi-step tasks, tasks requiring adaptation, long-running work.

The Cost of Choosing Wrong

Choosing an agent when a prompt suffices: unnecessary complexity, higher latency, higher cost, harder debugging.

Choosing a prompt when an agent is needed: brittle single-shot solutions that break on edge cases, require constant human intervention.

Decision Framework

Ask these questions:

1. Can this be done in one LLM call? → Use a prompt
2. Does it need real-time data or computation? → Add tools
3. Does it require multiple steps with dependencies? → Use a single agent
4. Do the steps benefit from parallel execution or specialization? → Use multi-agent

The Minimal Agent

The smallest useful agent has: a system prompt (role + behavior), one or two tools (file read, web search), and a loop (observe → reason → act → repeat).

Start minimal. Add capabilities only when the minimal version fails.

Anti-Pattern: Over-Engineering

The most common agent mistake is building a complex multi-agent system before proving a single agent can solve the core problem. Build the simplest thing that could work, then scale up based on real failure modes.