37. Agentic Workflows (Without Letting It Run Wild)

What is an Agent?

An "Agent" is just an LLM loop with tools.

while (goal_not_met):
  thought = model.think(context)
  action = model.decide_tool(thought)
  result = execute(action)
  context.append(result)

That's it. It's not magic. It's a `while` loop.

The Danger Zone

Agents are prone to:

• Loops: "I need to check the file." -> "I checked the file." -> "I need to check the file."
• Drift: Starting with "Fix the bug" and ending with "Rewrite the entire database layer."
• Cost: One user query can trigger 50 model calls ($$$).

Safe Patterns

We focus on Bounded Agents. Agents with a budget (max 5 steps), a strict toolset, and a "Supervisor" (another model or a human) that checks the plan before execution.

Where to go next

• 37.1 What an “agent” is in practice
• 37.2 Single-agent vs multi-agent designs
• 37.3 Planning, executing, and critiquing roles
• 37.4 Termination conditions and budgets
• 37.5 Auditability and replay