Chapter 6: LLM applications
Workflows vs agents: planning, memory and multi-agent orchestration
Many real jobs follow the same steps every time, and an agent that plans its own steps only adds choices that can go wrong. Interviewers ask "When would you not build an agent?" to see if you pick the simple design first. You will also learn how to split a job among several agents, and what an agent must remember.
What you will learn
- Choose between a fixed workflow and an autonomous agent
- Design short-term and long-term memory for an agent
- Compare orchestrator-worker and multi-agent patterns
Key terms
- workflow
- LLM calls wired together in a fixed order that the code sets, not the model
- planner
- an LLM call that splits a goal into subtasks and writes them as a plan
- subtask
- a smaller part of a larger job
- short-term memory
- what the agent keeps for the current session only
- long-term memory
- facts kept in a database that last across sessions
Prerequisites
Take An agent is an LLM in a loop with tools first. It comes right before this lesson in the course.
Where this lesson picks up: You built an agent: an LLM that thinks, calls a tool, observes the result and repeats these steps until done. That left a question open: should every LLM app be an agent?
What comes next
The workers send back their parts, so who checks that the result is good enough before the user sees it?
Next lesson: The ML system design skeleton
More lessons in this part
Take the lesson
Short cards, one idea at a time, with quick checks as you go.