What is an AI agent?

1. 1

   From answering to doing

   Suppose you ask why refunds increased. A ready-made aggregate may need only one summary. Scattered evidence may require queries, classification and calculations in an adaptive order. We will rehearse that investigation with fictional data.

   Products use “agent” differently. Here it means a configuration in which a model chooses subsequent steps and tools, contrasted with a predefined workflow. Real systems may combine both.

   A goal does not grant unlimited tool permissions. Investigating refunds is different from issuing them. Define readable aggregates, writable files and permitted communications separately; plans must stay inside that scope.

   Chatbot, workflow and agent compared

   Approach	Who decides the next step	Example
   Chatbot	A person, one question at a time	Paste in meeting notes and get a summary
   Workflow	A developer, in a fixed order	A message arrives → classify → draft a reply → staff review
   Agent	The LLM, step by step as results come in	To find why refunds rose, it queries data, calculates, then queries again

2. 2

   It loops: decide, use a tool, check

   Inside, an agent is usually simple. The LLM looks at the goal and picks a tool and its inputs; the app or service runs the tool and returns the result. The LLM reads the result, decides the next step, and repeats until the goal is met.

   

   Tools are capabilities wired in ahead of time: search, database queries, reading files, calculations, drafting an email. Because it sees a real result at every step, it can change course midway when the first plan turns out wrong.

   For a fictional calculation, a tool receives 30 and 20 and returns a 50% increase. Correct arithmetic does not establish that both counts cover comparable periods and definitions. Tool success is not evidence that its inputs were appropriate.

   An agent working on "Tell me whether refund requests rose this month compared with last month" (example)

   Step	The LLM decides	Tool used	Result
   1	Start with this month's messages	Query the support database (October)	412 messages
   2	Count only the refund requests	Tally by category	57 refund requests
   3	Last month's figure is needed to compare	Query the support database (September refunds)	41 refund requests
   4	Ready to answer	(no tool)	"57, up 16 from 41 last month"

   Check it

   • When you use an agent-based service, check that it shows a step-by-step log of which tool it used with which inputs. Without one, you can't tell at which step a wrong result crept in.

1. 3

   Good fits, and jobs it isn't ready for

   Agents suit work where you can't predict how many steps it will take or fix the order in advance. In exchange, every extra step adds time and cost, and a small mistake early on can carry through to the end.

   So start with the simplest approach that works. If one request does the job, use a chatbot; if the steps are always the same, a workflow is enough.

   Start with a workflow when the order is stable. A weekly sum of fixed columns has a clear deterministic solution; using AI does not require inventing a new plan every time. Adaptive selection can help when each finding changes the next useful question.

   Matching the approach to the job

   Kind of job	Best fit	Why
   Answering one question, summarizing, translating	Chatbot (one request)	One step, so no agent needed
   Repetitive work that always runs in the same order	Workflow	A fixed order is cheaper and more predictable
   Research or fixes where the number of steps is unknown	Agent	The next step depends on each result
   Hard-to-undo actions (payments, mass deletion)	A person does it, or a person must approve	A small misjudgment would be carried out as is

   I want to pick which of my team's tasks suit an agent. For each of the three below, tell me in a table whether it fits 'one request', 'a fixed workflow' or 'an agent', and why.
   1) Building the weekly sales report every Monday, always in the same format
   2) Finding out why customer complaints suddenly went up
   3) Translating a one-page contract from English

   These are editorial design judgments. A task with the same name may need a simpler method depending on evidence and permissions.

   The best fit for each task (example)

   Task	Best fit	Why
   Weekly sales report	Fixed workflow	The same data is gathered in the same order into the same format each week, so the steps can be set in advance
   Why complaints went up	Agent	There's no telling in advance which data to check, or how many times
   Contract translation	One request	It's a single step, but check your company's security rules before pasting in a contract

2. 4

   Settle these first: permissions, approval, stopping

   Because an agent carries out several steps on its own, the damage from a mistake, or from being tricked, can grow just as much. Run it thoroughly in a test environment first, and connect it to real work only after settling the four points below.

   Agents that read web pages, emails or documents can be fooled by instructions hidden inside them. This is called prompt injection, and attackers use it to steal data or trigger unwanted actions.

   “Until solved” is not a complete stop rule. For this paper exercise, stop after five tool attempts or two repeated query failures. Those are invented teaching limits, not universal safe defaults; real limits depend on cost and recoverability.

   A checklist before you connect an agent

   Settle	Why	Example
   Minimum permissions	Limits how far a mistake or an attack can reach	Allow reading only at first; block writing and deleting
   Human approval for hard-to-undo actions	Stops a small misjudgment from being carried out	A person confirms before any email is sent, payment made or file deleted
   Stopping conditions	Keeps it from looping on and running up costs	A maximum number of steps, a spending cap
   Treat what it reads as data only	So hidden instructions in web pages or emails can't steer it	Don't let it follow instructions found in outside content; review its log

   Check it

   • Start with a test account and fake data, and check the step-by-step log for what the agent did and why.
   • Test that the approval step really works, and that the agent actually stops when you don't approve.
3. 5

   Track failure, retries and approval as states

   Do not blindly repeat a failed tool call. Distinguish reads from state changes: repeating a lookup differs from repeating a refund. A delayed response does not establish that the first operation never happened.

   

   Use states such as investigating, insufficient evidence, awaiting review and complete. Silence in a review state is not approval. Identify the exact unresolved condition, and test that rejection actually blocks execution.

   Log necessary input summaries, tool names, results and output locations without indiscriminately retaining secrets or full customer records. A model’s later claim of success must agree with authoritative execution records. Open the output file to verify it exists.

This is a learning check, not a test you must pass. Use the explanations to understand the ideas.

1. Q1A refund-investigation agent has an issue-refund tool. Does the goal authorize using it?

   Write your thoughts (optional)

   Write your thoughts (optional)

   Your writing is not sent anywhere. It disappears when you leave or reload this page, so copy anything you want to keep.

   Read the answer and explanation

   No. Investigation and monetary action have different scopes. Stay within authorized reading, calculation and drafting; actual refunds require their own authority and process.

   I read and understood the explanation

   This is your own review, not an automatic score.

2. Q2The 20-to-30 comparison gives a 50% increase. Does it identify a cause?

   Write your thoughts (optional)

   Write your thoughts (optional)

   Your writing is not sent anywhere. It disappears when you leave or reload this page, so copy anything you want to keep.

   Read the answer and explanation

   No. Verify comparable periods and definitions; causal claims need additional evidence. These numbers are fictional practice data.

   I read and understood the explanation

   This is your own review, not an automatic score.

3. Q3Can an agent proceed because nobody answered an approval request?

   Write your thoughts (optional)

   Write your thoughts (optional)

   Your writing is not sent anywhere. It disappears when you leave or reload this page, so copy anything you want to keep.

   Read the answer and explanation

   Silence is not approval. Keep the action pending and state the needed decision. If a stop condition is reached, report the incomplete status accurately.

   I read and understood the explanation

   This is your own review, not an automatic score.

1. 6

   Try it: a five-step execution record

   Draw five rows and set a fictional goal: find two pieces of evidence for increased refunds and draft a report. Allowed tools are prepared aggregates, a calculator and draft writing; customer contact, refunds and production changes are outside the task.

   Record counts of 20 and 30, verify comparable periods, then calculate the increase. If the complaint-category table is missing, end with insufficient evidence instead of inventing the second reason.

   The draft says counts increased 50% while category-level causes remain unverified. All numbers are invented. A partial result with explicit evidence and missing data is useful. Rehearse two repeated query failures to verify the stop rule.