Page 122 - CyberSecurityEssentialsEnglish

Evaluate an agent at three levels. First, the quality of planning: did it divide the goal into
appropriate steps, or repeat work, or take a dangerous shortcut? Second, the quality of tool
use: did it choose the right tool, pass the right parameters, and check the result? Third,

operational control: are there limits on permissions, time, and cost, an approval step before
sensitive actions, a complete log, and an emergency stop?

A greater number of agents does not guarantee greater intelligence. Dividing roles among a

researcher, a critic, and a writer may improve organization, but it can repeat the same error
if all of them rely on the same source or if the critic is not independent. Compare a multi-
agent system with a simpler baseline. If it does not raise accuracy or reduce review time, it

is adding cost and complexity without benefit.

Begin by placing the agent in proposal mode: it prepares the draft or the plan of actions but
does not carry them out. Then allow low-impact actions, such as creating a draft inside a test
folder. Do not grant it external sending, deletion, purchasing, or the modification of

important data except with explicit approval and with limits that can be audited and
reversed.

⚠ Warning: permissions compound across a chain of tools

A small permission may expand through a chain of tools. The ability to read email, create
a file, and send a link may together expose information the user never intended to share.
Review the sum of the permissions, not only each tool on its own.

● Remember: Section summary
• Key idea: Compare within the right category, and read the license rather than the
marketing word.

• Practical decision: For agents, start in proposal mode and expand permissions only
against evidence.
• Mistake to avoid: Treating open as a single meaning, or assuming that more agents
means more reliability.

3.5 How Do We Improve the Fit Between System and Task?

◆ Key idea: there is no mandatory ladder.
There is no fixed sequence that begins with prompting, then retrieval, then fine-tuning.

Each intervention addresses a different kind of failure, and more than one can be
combined in a single solution. Diagnose the cause first.

121