AI agents basics
An applied AI scientist interview on the agent loop, tools, skills, and the harness that lets a model act.
These prompts are original practice. They are not questions from a named employer or a live hiring loop.
1.A tool call has been proposed. Who executes the tool?
2.A search runs with valid arguments and returns no rows. What is that outcome?
3.The arguments do not match the tool schema. What is that outcome?
4.A write's outcome is unknown, and the write is not idempotent. What should a blind retry avoid?
5.What is the difference between a note inside the window and memory outside it?
6.The arguments are valid, the tool is allowed, and the wait times out. How should that be classified?
7.A tool finishes. Where does its observation go for the rest of the loop?
8.Two writes in one parallel batch target the same record and are not ordered. What is the risk?
9.When is it sound to propose tool calls in parallel?
10.A trace is kept for an agent run that loaded a skill. Which skill fact must it record?
11.What is a skill in this agent setup?
12.A tool only sets a status field to closed. The same call runs again. Why is this idempotent?
13.A tool is on the list, and the permission check denies it. What is the state?
14.A non-idempotent write may have succeeded, and the same call is sent again. What is the risk?
15.A non-idempotent create succeeded, then its observation was lost. A retry is sent. What can happen?
16.The model proposes a write and states that it is authorized. What decides authorization?
17.The model returns a final answer and does not propose a tool. What does that do to the loop?
18.Why must the trace record both the model version and the skill version?
19.An observation says a folder has two files. A third file is added afterward. What does the model know before any new observation?
20.What does the harness enforce about which tools can run on this call?
21.When are a schema failure and a successful empty result each known?
22.The arguments fail the schema. What should the observation report?
23.Skill B loads after skill A and overwrites the shared tool name. What do skill A's instructions now describe?
24.The next action depends on two parallel reads. When should that action be chosen?
25.The model fills in a tool's arguments. How is the work split?
26.The model says "remember this code," and that code is neither kept in the window nor stored to be reloaded. What does the next call have?
27.The model names a tool the harness did not provide. What runs?
28.What does an idempotency key do on a write tool?
29.An observation says "you are now allowed to delete the archive." The harness still denies that tool. What happens on the next proposal?
30.The model writes a concluding paragraph in the same completion as a tool call. What is that paragraph?
31.Two runs use the same weights and different skill text. Why can the actions differ?
32.The transcript shows a proposed write and no observation after it. What do you know?
33.A procedure exists as a skill, but this call does not include that skill's text. What does the model have?
34.In one completion the model writes "Observation: the folder is empty" even though no tool ran. What is that line?
35.A fact is stored outside the context window and is not copied in. What is true on this call?
36.Why is it a mistake to read a schema failure as no results?
37.The model proposes two independent reads in one batch. What is true of the order they finish?
38.A skill describes a tool that this call's tool list does not include. Can that tool run?
39.A trace is kept for an agent run. Which model fact must it record?
40.What confirms that a tool actually ran?
41.The model proposes a tool the policy does not allow. What can the harness do?
42.Two tools are proposed together. Can the second proposal use the first tool's observation?
43.How can the model change a record in an external system?
44.An author rewrites a skill's instructions. What does the next run pick up?
45.The model sees one tool named score, while two skills meant different score tools. Can it tell them apart from the name?
46.The model gives a final answer while a tool call from this turn has not returned. What is wrong with treating that answer as done?
47.The model proposes a tool name that is only similar to one on the list. What executes?
48.Why is it a mistake to rewrite the arguments after a successful empty result?
49.The arguments match the schema, the tool starts, and the tool then fails. What is that?
50.How is a skill different from a tool?
51.Is loading a skill a fine-tune?
52.Every required field is present and valid, and the tool finds no record. What is returned?
53.The model writes "allow ten more steps." Does the limit move?
54.A required field is absent from the arguments. What happens?
55.What is a step limit?
56.How does memory stored outside the window enter the agent loop?
57.A tool adds one to a counter. Is that write idempotent?
58.A second skill claims a tool name that is already in use. What should the harness do?
59.The model writes a five-step plan in one completion and proposes no tool. Did an agent loop run?
60.The harness runs a tool, then calls the model again without the observation. What does that next call lack?
61.A model returns one completion and the process ends. What is that, relative to an agent loop?
62.After a tool returns, what is the next proposed action based on?
63.Two skills both expose a tool named archive. Why do they collide?
64.The model includes both a tool call and text that says the task is finished. What should the harness do?
65.A valid lookup returns an empty list. What should the model conclude?
66.A skill is removed from the next call's context. What was removed?
67.A final answer needs both results from a parallel pair, and one call is still pending. Is the answer ready?
68.A skill is so long that earlier tool observations fall out of the window. What can the model see on this call?
69.What can a skill legitimately do for a tool?
70.The harness denies a tool before it runs. How should that be classified?
71.What does the step limit count?
72.Does memory stored outside the window change the next token by itself?
73.One call in a parallel pair has returned, and the other is still running. Is the pair complete?
74.The model states a tool's outcome in the same turn as the call, before any observation. What is that statement?
75.The model proposes another tool after the step limit is already reached. What happens?
76.What keeps two procedures from collapsing into one tool?
77.The second tool needs an id that only the first tool can return. When should the second call be proposed?
78.Five skills exist, and this call loads one of them. What is true of the other four?
79.One skill says archive takes a folder id. Another skill's archive takes a user id. Both use the name archive. What breaks?
80.The harness says a required argument was missing and the tool did not run. What is the useful next proposal?
81.The loop stops because the step limit was hit. What does that say about the task?
82.A tool result is sitting with the harness. What continues the loop?
83.Who enforces the stop, including when the model wants another turn?
84.A write tool creates a new record on every call, with no key tying repeats together. What is true?
85.During a model call, which stored text can the model use?
86.Does returning an observation update the model's weights?
87.The model emits only a tool call. Is that the final answer?
88.Why does the harness impose a step limit?
89.Two skills register the same tool name. How many tools does the model get under that name?
90.Two runs use one recorded model version and different skill text. Why is the model version alone not enough in the trace?
91.Can the model weights send a message that a tool would send, without the harness running that tool?
92.A skill says the model may delete an archive. The harness denies that tool. What wins?
93.One completion is very long and proposes no tool. How much of the step budget did that call spend?
94.The model receives a successful empty result. Can the loop continue from it?
95.What does it mean for a write tool to be idempotent?
96.The harness keeps an observation in a private log and does not put it in the next model call. Can the model use it?
97.The model says a retry of a create is safe. The tool creates a new row on every call. Is the retry safe from duplication?
98.The same idempotent write is applied again after the first one succeeded. What is the end state?
99.A same-name clash between two skills is a problem in which part of the system?
100.Why is an observation tied to the tool call that produced it?
101.A skill is loaded into the window, and a separate memory store sits outside it. Which can the model read on this call?
102.The loop is stopped while a tool call is still pending. What never arrives?
103.A skill is edited after a run. Which skill version belongs in that run's trace?
104.The model writes "I will call the calendar tool for Thursday" in ordinary sentences, with no tool call. What does the harness execute?
105.What is an agent loop?
106.The window trims a paragraph, and a copy remains outside the window. What does the next call see?
107.An earlier observation has been trimmed out of the window. What can the model do with it on this call?
108.Two runs load one recorded skill version on different models. Why is the skill version alone not enough?