What Are AI Agents?

The Agent Spectrum

Not all AI systems are created equal. Understanding where different AI products sit on the autonomy spectrum is critical for product decisions.

Chatbots: Reactive Responders

Traditional chatbots (even LLM-powered ones) are reactive. They respond to a single prompt, generate output, and wait for the next input. Think ChatGPT in its simplest form — you ask, it answers.

Key characteristics:

• Single turn or multi-turn conversation
• No ability to take actions in the real world
• No persistent goals or task tracking
• Output is always text (or structured text)

Copilots: Augmented Human Workflows

Copilots sit alongside humans and augment their work. GitHub Copilot is the canonical example — it suggests code as you type, but you're still driving.

Key characteristics:

• Embedded in an existing workflow
• Suggests actions but doesn't execute autonomously
• Human makes final decisions
• Reduces friction in existing processes

Agents: Autonomous Goal Pursuers

AI Agents are systems that can autonomously pursue goals by planning, using tools, and iterating on their approach. They're the most capable — and most complex — AI pattern.

Key characteristics:

• Given a goal, not just a prompt
• Can plan multi-step approaches
• Use tools (APIs, databases, code execution)
• Can observe results and adjust strategy
• May involve human-in-the-loop for critical decisions

The Agent Loop

Every agent follows a core loop:

1. Observe — Receive input or observe environment state
2. Think — Reason about what to do next (LLM inference)
3. Act — Execute an action using a tool
4. Reflect — Evaluate the result and decide next steps

This loop continues until the goal is achieved or the agent determines it cannot proceed.

When to Build an Agent vs a Chatbot

Ask yourself these questions:

• Does the task require multiple steps? If yes, lean toward agents.
• Does it need real-world actions? (API calls, file operations, data queries) — agents.
• Is the task well-defined with clear success criteria? — agents work best here.
• Is the value in conversation itself? — chatbot might be sufficient.
• Do you need predictable, low-latency responses? — chatbots are simpler and faster.

Key Takeaways

1. Agents are a superset — every agent can chat, but not every chatbot can agent.
2. Start simple — build a chatbot first, then add agentic capabilities incrementally.
3. Tools are what make agents powerful — the LLM is the brain, tools are the hands.
4. Human-in-the-loop is a feature, not a limitation — it builds trust and catches errors.