Chapter 5: The Anatomy of an Agent Loop

At its core, every agent follows a simple loop:

1. PERCEIVE — Take in information (user input, tool results, context)
2. REASON — Decide what to do next (analyse, plan, choose action)
3. ACT — Execute the chosen action (call a tool, generate output)
4. OBSERVE — Check the result (did it work? what changed?)
5. REPEAT — Continue until the goal is achieved or escalation is needed

This is sometimes called the ReAct pattern (Reason + Act) — the agent explicitly reasons about what to do before doing it.

What Comes Built-in vs What You Configure

Modern LLMs have ReAct-like reasoning baked in at a foundational level. When you give them tools, they naturally:

• Reason about what the user is asking
• Decide whether to use a tool or respond directly
• Observe tool results and continue reasoning

You don't have to explicitly instruct "think step by step, then act, then observe" — that's how the models work out of the box when tools are available.

The Platform's Role

The platform handles the orchestration loop:

• Sending the user's message to the model
• Executing tool calls when the model requests them
• Passing results back to the model
• Continuing until the model produces a final response

Your Instructions Shape How the Agent Reasons

Your instructions (DNA) shape how the agent reasons, not whether it reasons:

• What persona and tone to use
• What domain knowledge to apply
• When to use which tools
• What guardrails to follow
• When to escalate or defer