Twelve questions on the lesson AI agents, each answer with its reason. The questions come from the pool of the module and are in the format of its exam.
You ask a chatbot to fix a script that stops with an error. It suggests a change, you run the script yourself and paste the new error message back. What changes if you give the same task to an agent?
It uses a larger language model that makes fewer mistakes when programming.
A chatbot and an agent can run on the same model. The difference is not the model but what is built around it.
It runs the script itself, reads the error and decides on the next change until the script works.
Correct: the agent takes over the steps you take with a chatbot. It calls tools, observes the result and carries on until the goal is reached.
It solves the task in one step, because it plans the whole solution before writing anything.
An agent does not work in one step but in a loop: it acts, looks at the result and then decides again. A plan helps it but does not replace the loop.
It asks for permission before each suggestion but otherwise works like the chatbot.
An agent may ask for permission, but that is not what sets it apart from a chatbot. What matters is that it carries out the steps itself instead of leaving them to you.
An agent reads the file fields.csv. What exactly does the language model contribute to this step?
It opens the file on the disk and loads its content into its memory.
A language model cannot open a file. It only produces text; everything that happens in the world is carried out by a program around the model.
It writes text that names the tool and its arguments; a program around it reads the file and returns the content.
Correct: the tool call is just text, like the TOOL card in the game. The harness carries it out and puts the result into the context.
It sends a request to the operating system, which checks the permissions and delivers the file.
The model does not talk to the operating system. The harness makes that request after it has recognised the model's text as a tool call.
It recalls the file from training, because tables like this occur there frequently.
A file on your computer cannot have been part of the training. Whatever the model knows about it is in the context only if a tool has read it.
Which of these tasks is better solved with a fixed workflow than with an agent?
Every morning an LLM reads the new delivery notes, a program matches them with the orders and flags differences.
Correct: the steps are the same every time and can be fixed in advance. That an LLM does one of them does not make it an agent.
A dairy's main supplier of packaging has gone bankrupt; find alternatives, compare offers and draft a recommendation.
Nobody knows the necessary steps in advance here: which sources, which suppliers, which follow-up questions. This is exactly what an agent is for.
Find out why the yield on one field dropped last year, using all data and sources available.
A search for causes branches with every finding; which step comes next depends on the result of the previous one. That points to an agent.
Build a small web app showing the bookings of the farm machinery, from a short description.
When building an app, the next step follows from the last one: read the error, change the code, test again. That is an agent's loop.
Which of these jobs does the harness do rather than the language model?
carry out the tool calls and return the results
Correct: the model only writes the call down; the harness carries it out, as in the game, where the Harness carries every card to the tool.
decide which tool to call next
The model makes this decision. The harness checks and carries out what the model writes, but it does not choose the next step.
ask the user for permission before a command runs
Correct: asking is a rule in the program around the model; the model can neither trigger nor bypass it.
manage the context window, for example by summarising when it fills up
Correct: the harness decides what goes into the context window and what gets shortened when space runs out; the model only sees the result.
write the plan for the task
The model writes the plan, through a tool like any other. The harness stores it and shows it again, but the content comes from the model.
After two hours of work an agent suddenly repeats a step it had finished long ago. Which explanation fits best?
The model learned from the previous steps and adjusted its weights, which overwrote the earlier result.
A model's weights do not change while it is being used. Whatever it knows during a task sits only in its context window.
The context window was full; earlier results were summarised or removed, so the step is no longer in it.
Correct: what is no longer in the context, the model does not know. This is why agents write plans and notes into files they can read again.
A tool reported an error, and the harness repeats failed calls on its own.
An error would come back as a result, and the model would react to it. The harness does not repeat a step that was finished long ago on its own.
The language model only ever keeps the last tool result, so every earlier step is lost after one round.
Then the agent could not have worked for more than a few minutes. The context window holds every step until it is full.
An agent is asked to research 40 milestones in the history of AI and starts eight sub-agents for it. What is the main advantage?
Each sub-agent has a fresh context window, so the many sources do not fill the main agent's context, and they work at the same time.
Correct: the main agent only receives the results, not every page that was read. Its context stays free for the actual task.
Sub-agents use a smaller model that specialises in searching the web.
A sub-agent may use a different model, but it does not have to and usually uses the same one. The advantage lies in the separate context and in working in parallel.
Sub-agents may access files the main agent has no permission for.
Sub-agents run in the same harness with the same or narrower permissions. They may not do more than the agent that starts them.
Several sub-agents vote on every answer, which rules out invented facts.
Each sub-agent researches a part of its own; there is no vote. Invented facts remain possible and have to be checked.
You use Claude Code on your laptop, and it creates files in your project folder. Where does the language model run?
On your laptop, inside the program you installed.
The harness runs on the laptop, not the model; a model of this size does not fit on a laptop.
On a server of the provider; only text travels back and forth, while the harness and the files stay on your laptop.
Correct: the harness sends the context to the model in the cloud and carries out its tool calls locally. That is why the files appear on your machine.
Partly on the laptop and partly in the cloud: small steps locally, large ones on the server.
The model is not split by the size of a step. Every decision is made by the same model on the server; only the tools run locally.
In the browser, which downloads the model at the start of the session.
Claude Code does not run in the browser, and the model is not downloaded. Running a model locally is possible, but it is a separate choice with smaller models.
Your agent works with your user account, and you have allowed it to run commands without asking. What can it do?
delete files in your personal folder
Correct: anything your account may do, the agent may do as well, deleting included.
publish a website through your GitHub account
Correct: if your account is signed in on the computer, a command from the agent can publish in your name.
read files on a colleague's laptop on the same network that your account has no access to
The agent has exactly your permissions, no more. What your account cannot reach, the agent cannot reach either.
read every file your account can read, private documents included
Correct: the agent does not distinguish between project files and private documents; it can access everything you can.
change the model's weights so that it remembers your project next week
The weights belong to the provider and do not change through use. An agent can only remember through files it reads again.
Two agents use exactly the same language model but different harnesses. Which statement is correct?
They behave the same, because all decisions are made in the model.
The model decides, but only about what the harness puts in front of it. Other tools, instructions and permissions lead to other decisions.
They can behave differently, because the harness determines which tools, instructions, permissions and context the model gets.
Correct: an agent is an LLM plus a harness. The same model in a different harness is a different agent.
They differ only in speed, because the harness sets how fast the model computes.
How fast the model computes is determined by the provider's server. The harness changes what the model sees and may do, and with it the behaviour.
They give the same answers in a different layout, because the harness only formats the output.
The harness does more than format; it runs the loop, the tools and the context. Which answers come about at all depends on that.
An agent is asked to work out how much nitrogen may still be applied to a field. Its web search returns a current page and an archived page from 2009 with a different requirement, and it calculates with the 2009 value. What does this show?
The search tool made a mistake; a better one would only return current pages.
The search did what it was asked: it found matching pages. Old sources exist on the real web too, and an agent has to deal with them.
The model works only with what is in its context; whether it notices an outdated source depends on whether it reads and checks the result carefully.
Correct: the year was in the search result. An agent is only as good as the results that reach it and the way it handles them; in the game "Be the Agent" this is exactly the trap of Task 2.
The harness should have filtered out the old page, because it checks every result for correctness.
The harness checks whether a call is valid, not whether a result is correct in substance. Judging that remains the model's job.
A larger model would have known the current value from training, so the mistake is down to the size of the model alone.
Training knowledge has a cut-off date and may itself be outdated. That is exactly why an agent searches for sources and has to judge how old they are.
Which sequence describes the loop at the core of every agent?
plan, carry out all steps, deliver a report
This lacks the observing between steps. An agent looks at every result before it chooses the next step.
prompt, answer, user feedback, new prompt
That is a conversation with a chatbot: the human takes the next step. With an agent the loop closes without the human.
think, act, observe, and again until the goal is reached
Correct: in research this pattern is called ReAct (Yao et al., 2022); in the game it is the THINK, TOOL and RESULT cards.
search, read, summarise, and stop after one answer
This is a fixed sequence with a single pass. An agent has no fixed number of passes; it stops when the goal is reached.
METR reports a "time horizon" of several hours for current models. What does it measure?
How long a model can work before its context window is full.
The measure has nothing to do with the context window. It measures the length of tasks, not how long the model runs.
The length of tasks, measured in the time a human needs, that a model completes on its own in half of the cases.
Correct: a horizon of four hours means that tasks a human needs four hours for succeed in about half of the cases.
How many hours of human work a model saves on average per day.
METR does not measure saved working time. Whether the time horizon translates into savings is a separate, open question.
How long the model needs for a task that a human completes in one hour.
What is measured is not the model's computing time but how long the tasks it still manages may be.