Fawzooz
عربيCollaborate
Your journey
  1. 1
  2. 2
  3. 3
  4. 4
  5. 5
  6. 6
0 of 34 units

You are in Degree 3 · Practical masteryunit 3 of 6Ahead of you: A real piece of professional work, and the page that shows you checked it.

Degree 3 · Unit 3.3

Agents: a system that acts

An agent is not a cleverer model. It is a model that has been given four things: a goal instead of a question, tools it can actually operate (search, email, files, databases), a memory in which it accumulates what it has already done, and a loop in which it keeps trying until it reaches the goal or stops.

This shift — from answering to acting — is the most important thing to have happened in this field in years. And with it, the question moved from "is the answer right?" to something a great deal heavier: "what can this system actually do if it gets something wrong?"

FIG. 14 — The agent loop
1
It plans
It breaks the goal into steps
2
It picks a tool
Which of the available tools completes the step
3
It acts
It runs the tool and reads the result
4
It evaluates and repeats
Has it reached the goal? If not — back to the first step
↻

The danger here is not a single error but the way errors accumulate around the loop: a third step built on a conclusion that was already wrong in the first. And an agent will not doubt itself unless you have made it do so.

The delegation standard

Do not delegate by instinct. Ask two questions of every task: can this action be undone? and would the error show up immediately? The answers to those two place the task somewhere in the matrix below.

FIG. 15 — The delegation matrix
Reversible · error shows immediatelyDelegate entirelySearch, summarising, drafts, sorting files
Reversible · error shows lateDelegate with periodic reviewClassification, data analysis, internal follow-up
Irreversible · error shows immediatelyApproval before every actionSending a client email, publishing, deleting
Irreversible · error shows lateNever delegateA financial transfer, a contract, a decision touching a person

Five constraints without which no agent runs

1Least privilege: give it the narrowest access that will complete the task, rather than handing it your own.

Think, act, observe

The difference between a request and a goal: you set the end, and the agent derives the steps and reaches for its tools — a browser, a calculator, a database, a translator, another model.

FIG. B16 — The loop and the toolbox
ThinkActObserveThe intelligent agentOne goal · many tools
TOOLBOX
Web browserCalculatorDatabaseTranslatorAnother model

The loop can turn for a long time, which is why an agent needs a cost ceiling, a step limit and a stopping point where a person reviews it. One goal, many tools — and that is what makes the five constraints above a condition rather than an option.

Do this

  1. 1 — On paper. Place ten tasks from your work in the matrix. Look at the last square: these are the tasks that will stay human however far the technology advances.

  2. 2 — On the tool. Run an agent on a task from the first square, and read its full step log once it has finished. How many steps were redundant? And where did it nearly go astray?

  3. 3 — In your field. Write the three human stopping points you will not give up, however good the agent is.

Where to after this unit? You have run a general system. The next unit makes it yours: an assistant that knows your own knowledge.