Lesson 1 / 25

The Spectrum of Autonomy

Place a system on the scale from plain assistant to autonomous agent and choose the right level.

Not a yes/no label

"Agent" is a spectrum. At one end is a plain assistant that only answers. Next comes an assistant with one or two tools. Then a workflow where your code fixes the steps. At the far end is an autonomous agent that chooses its own steps and may run for many turns. More autonomy gives more power, and also more cost, risk and unpredictability.

From answering to acting

An agent perceives a goal, reasons about what to do, acts through tools and learns from the result.

Four stages: goal, reason, act, observe.
Figure 1.1 — Goal, reason, act and observe.

A quick decision guide

Use the lowest level that solves the problem. Move right only when a simpler design measurably falls short.

Answer from knowledge only        -> plain assistant
Need 1-2 lookups or actions        -> assistant + tools
Steps are known in advance         -> workflow (code controls flow)
Steps depend on what is found      -> agent loop
Long-running, many tools, risky    -> agent + approvals + budgets

Cab, bus and self-drive

A bus follows a fixed route (workflow). A cab driver picks the route given a destination (agent). Hiring a cab for a bus route costs more, and trusting a driver you cannot watch needs seat-belts and rules.

Quick check: When should you prefer a fixed workflow over a full agent?

  • Never
  • When you want more risk
  • When the steps are known in advance
  • When tools do not exist
Answer

When the steps are known in advance — Known steps do not need the model to decide them, so a workflow is cheaper and more predictable.