Stage 3 · Expert · L7
Agents
A model in a loop with tools is a different kind of software.
12 lessons · 119 minSteady
About this chapter
An agent is a loop: think, call a tool, read the result, repeat until done. That small change turns a text generator into something that can act, and it brings every problem that comes with acting. This chapter covers tool design, memory, permissions and cost control, because those are what decide whether an agent is useful or a liability. You will also learn to evaluate agents on tasks rather than on transcripts that read well.
What you will be able to do
- 19 min
The Loop
Draw the think-act-observe loop and name its stopping conditions.
- 210 min
Tools Are Functions With Descriptions
Write a tool definition the model can call correctly on the first try.
- 310 min
Tool Output Is Data, Not Orders
Stop an agent from acting on text it read: gate on where the text came from, and never hand one agent all three legs of the trifecta.
- 410 min
Planning
Compare upfront plans with step-by-step decisions and pick one per task.
- 510 min
Memory Between Steps
Design what an agent carries forward and what it recomputes.
- 610 min
More Than One Agent
Say when splitting into several agents helps and when it just adds cost.
- 79 min
Tool Protocols
Describe how a shared tool protocol lets one agent reach many systems.
- 811 min
Guardrails and Permission Models
Classify actions by reversibility and require approval for the risky ones.
- 910 min
Loops, Hallucinated Tools and Stalls
Detect the common agent failure modes and add a guard for each.
- 1011 min
Evaluating Agents
Build a task-based evaluation with a checkable end state.
- 119 min
Keeping the Bill Sane
Bound steps, tokens and retries, and log spend per run.
- 1210 min
What They Can Do Today
Quote the published numbers on agent success, then and since, and pick tasks that fit them.