◢ Chapter 01 · Foundations

Before "agent" meant anything, there was a spectrum.

Most confusion about agents comes from not knowing where on the autonomy spectrum you actually are. Pick the right slot — don't reach for multi-agent when a workflow will do.

The autonomy spectrum →
LLM picks its own steps

Single Agent

The LLM decides which tools to call and in what order, in a loop, until it decides it's done. More flexible, less predictable.

Cost
$$$
Latency
5-30s
Predictability
★★★
◢ PM insight
Climb the spectrum reluctantly. Every step right multiplies cost, latency, and surprise. "Agentic" is a tool, not a virtue. The LinkedIn Writer in Chapter 5 is technically multi-agent — but the interior of each agent is mostly a tight workflow. That's the trick: outer autonomy, inner determinism.
◇ See also· Autonomy spectrum
→ PM Pattern: Cheap-First Cascade→ Briefing flow: how it picks its slot
◢ The atom of agents

Every agent is built on one idea: the augmented LLM.

A naked LLM is just a function: text → text. It becomes an agent when you wrap it with three things — retrieval, tools, and memory — so it can reach beyond its weights.

Retrieval (RAG)vector DB · documents · search index
Toolsweb search · code exec · APIs · DB writes
Memoryconversation · scratchpad · long-term store
LLM
generates · plans · decides
↓output
Action / Response
Retrieval
Pull relevant context into the prompt at runtime. Vector search, BM25, hybrid, knowledge graphs. Without it, the model is stuck at its training cutoff.
Tools
The model emits a structured "I want to call X" message; your code runs X and feeds the result back. This is how agents act on the world.
Memory
Short-term (conversation), working (scratchpad), long-term (DB/embeddings). Memory is what makes an agent feel continuous.
◇ See also· Retrieval · Tools · Memory
→ PM Pattern: Memory Write-Approval→ PM Pattern: Voice-Anchored Generation→ Briefing flow: where memory plugs in
◢ When to choose what

Pick the smallest thing that works.

Can one prompt solve it?
→ Yes
Use a Simple Prompt.
Are the steps known in advance?
→ Yes
Use a Workflow (chain / route / parallelize).
Does the task require dynamic tool use & branching?
→ Yes
Use a Single Agent (ReAct loop + tools).
Are there genuinely distinct expert roles producing one deliverable?
→ Yes
Use Multi-Agent (orchestrator + specialists + evaluator).
Next chapter: the patterns library →