Project 42Start learning
Module 2 of 5 · 38 min

Classify Who Chooses the Next Step

Distinguish model calls, deterministic workflows, agents, and agentic systems by executable control flow and feedback.

Core concept

By the end

You will be able to

  • Trace code-directed and model-directed transitions separately.
  • Identify the loop, feedback, state, stop conditions, and controller.
  • Explain why tool-call output does not by itself create an agent.
  • Choose the least complex accurate classification.
01

A model call produces an output

A model can generate an answer, plan, or tool-call proposal without controlling an executable loop. The caller still decides what happens next.

Tool-call syntax describes a requested action. An authorized runtime must validate arguments, resolve the target, execute or refuse the call, and return evidence.

02

A deterministic workflow follows predefined transitions

Code or configuration selects the next step in a deterministic workflow. A model may generate or classify inside a step while the workflow remains the controller.

Deterministic does not mean perfectly predictable output. It means the executable path and transition authority are defined outside the model.

03

An agent chooses among permitted next steps

A bounded agent pursues a goal through a loop: observe state, choose an allowed action, receive feedback, update state, and continue or stop. External controls still own identity, authority, budgets, verification, and terminal outcomes.

An agentic system is the complete operating boundary around one or more agents: harness, tools, state, policy, telemetry, evaluation, approvals, recovery, and accountable people.

04

Multiple calls are not multiple agents

A multi-agent system has separately instructed or controlled agent units that coordinate, delegate, or hand off work through explicit contracts.

Parallel requests, debate prompts, or multiple models do not prove multi-agent behavior unless separate responsibilities and coordination are observable.

Practice activity

Trace three equivalent designs

  1. Represent the same task as one model call, a deterministic workflow, and a bounded agent loop.
  2. Mark every transition as user-directed, code-directed, model-directed, or unknown.
  3. Identify tool execution, feedback, state update, verification, and terminal control.
  4. Choose the least complex design that meets the declared requirement and justify the classification.

What to produce

  • Three comparable control-flow traces with transition ownership.
  • A classification decision tied to observable signals and explicit unknowns.

Reflect before continuing

Which transition actually required adaptation rather than predefined code?

Evidence

Sources and verification

Knowledge check

Make it stick.

Pass at 80%

Choose the strongest answer for each question. Your attempts become part of your device-local transcript.

01What most directly distinguishes an agent loop from a deterministic workflow?
02A model returns a function name and arguments. What has happened?
03A fixed pipeline calls a model in step two. How should it be classified?
04What proves a multi-agent system?
05What is the preferred design when a fixed workflow meets every requirement?