Page 213 - CyberSecurityEssentialsEnglish

Chapter Eight on institutional oversight, because the system is no longer following your
sequence.

FIGURE 5-5: FROM WORKFLOW TO AGENT, FIVE LEVELS OF AUTONOMY

What to notice: The words are often used loosely, but the difference between them is a
difference in who decides the next step, and therefore in which controls apply.

Source: Prepared by the author.

● A decision rule for the vocabulary
If the system decides the next step, calls tools dynamically, observes the result, and may

re-plan, treat it as an agent and govern it as one.
If the sequence is defined in advance, treat it as a workflow, however capable the model
inside it may be.
The label matters because it determines whether you owe it a stopping point, an
approval gate, and an action log.

5.4.3 What a Production Agent Is Made Of?
An agent in production is not a prompt with ambitions. It is a small system built around a
model, and it fails in the place where a component is missing rather than where the model

is weak. Six parts are worth naming, because each one is a thing you can inspect, test, and
hold someone responsible for.

The model is the reasoning core, a language or multimodal model. The tools are what it may
call: search, code, files, databases, or an interface to another system. The memory is the

state it keeps, both within a session and, sometimes, between them. The planning is how it
breaks a goal into steps and what it does when a step fails. The guardrails are the controls
placed around it: permissions, validation, allowlists, and approval gates. The observability is

your ability to see what it actually did: traces, logs, cost, latency, and recorded failures.

212