AI agents basics
An AI/ML engineer 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.What is an agent loop?
2.A model answers in one completion and never calls a tool. What kind of run is that?
3.Tools are listed on the call, and the model answers once without using any of them. What ran?
4.One tool call returns, and the model then gives a final answer. What is that sequence?
5.Between steps, who executes a tool and produces the observation?
6.How does the next model call know what a tool returned?
7.A turn contains prose and a tool call together. When can later prose depend on that tool's result?
8.What does the loop do with an observation, relative to the call that requested it?
9.What is a tool call?
10.The model writes a sentence telling itself to look up order 18 and email the buyer, and it emits no structured call. What runs?
11.The model has streamed only part of an argument string. What should the harness run?
12.The model calls a name that was never registered. What should the harness do?
13.What is an observation?
14.What does the tool schema tell the model?
15.How does a required argument differ from an optional one?
16.The schema says status must be one of open, paid, or void. The model sends the value done. What fails?
17.A developer changes a tool's argument list in the schema and does not train the model. What does the next call see?
18.The schema says a tool previews a message, and the implementation sends it. Before the result returns, what has the model been shown?
19.A tool's schema is registered, and the tool server is down when the call is made. What is true?
20.Who chooses the argument values in a tool call?
21.When does the harness validate arguments against the schema?
22.The user said order 18. The model sends order 19, and 19 matches the schema's type. What does validation do with 19?
23.A required field is absent and the other fields match. What should happen?
24.The schema requires a string id, and the model sends a number. Under a strict schema check, what happens?
25.The arguments match the schema, and the tool then reports that the record does not exist. What is that report?
26.The model receives an error that a required field was missing. How can it correct the call?
27.A tool fails while doing the work. What should the model receive?
28.The tool finishes and the harness appends nothing. What can the model conclude?
29.A lookup succeeds and finds zero rows. How should that observation differ from a crash?
30.What does idempotent mean for a tool?
31.A charge tool stores nothing about earlier requests. What is a second identical charge call?
32.What lets a retry of a charge or a send avoid a second effect?
33.The model passes an idempotency key. When does that key stop a second charge?
34.Where does an automatic retry of a failed tool live?
35.Two identical charge calls are requested together, and the tool does not dedupe. What can running them together produce?
36.A send call times out, and the tool does not dedupe. What does the timeout establish?
37.The model emits several tool calls in one turn, and the harness runs them. What comes back?
38.The model requests several tool calls in one turn. Who decides whether they run at the same time?
39.Parallel tools finish in a different order than the model requested. How is each result identified?
40.A later tool needs an id that an earlier tool returns. In an agent loop, when can the model put that id into an argument?
41.Two separate tool calls both succeed. What makes them one transaction?
42.A list tool accepts many ids and returns one payload. How does that differ from parallel tool calls?
43.A list tool returns the first page plus a continuation token. What does the model have so far?
44.Which actions can run on this turn?
45.What does an allowlist of tools change?
46.A tool server advertises twenty tools. The harness exposes three of them to the model. How many can the model call?
47.What is a permission in an agent harness?
48.The prompt says not to call a tool. The tool is on the allowlist, and the model calls it anyway. What decides whether it runs?
49.The harness requires a person to confirm a loaded tool. What does that requirement insert?
50.The model writes the word approved in its reasoning and then calls a tool that requires a person. What is that word to the harness?
51.A person denies a tool that was waiting for confirmation. What should the loop receive?
52.What does on whose behalf select when a tool runs?
53.A call matches the schema, and the permission for this run refuses it. What runs?
54.The schema offers a write, and the attached token may only read. The call is well formed. What can the service do?
55.What bounds the authority of a run?
56.A page returned in an observation says to ignore the instructions and to call a send tool this run is not allowed to call. What does that page change in the harness?
57.Untrusted text in an observation urges a call to a tool this run is already allowed to use. What is that tool's authority after the text arrives?
58.What do a permission check, secret redaction, and a call log share?
59.The harness removes a secret before it builds the observation. What does the next call contain about that secret?
60.A tool uses an API key stored by the harness. The key is left out of the prompt and the observation. What does the model see of the key?
61.The model never received the API key. How can the tool still make the authorized request?
62.What is a skill?
63.What does loading a skill change?
64.An instructions-only skill is loaded. What executable name does that skill add?
65.A loaded skill includes a tool. What can the model do with that tool on a call where the harness exposes it?
66.Where can the skills and tools available on a call be announced to the model?
67.A skill was never loaded, and the model names it in a call. What can run?
68.The instructions describe a tool the harness never registered. What does that description make executable?
69.Tools A and B are both loaded and permitted. The prompt says to prefer A, and the model calls B. What can the harness do with B?
70.How is loading a skill different from handing a subtask to another agent?
71.A tool is absent from the first model call. The harness includes it on a later call. Which call can use the tool?
72.Which conditions can end an agent loop?
73.A loop has no final-answer rule, no step limit, and no tool that ends the task. What is missing?
74.What does a step limit do?
75.The step limit is reached before the model produces a final answer. What kind of ending is that?
76.The harness accepts a tool whose job is to end the task. What does accepting that call do?
77.The stop condition is a final answer. What is that answer?
78.The model emits a preamble and a structured tool call. The stop condition is a final answer. What does the harness execute?
79.The stop condition has already been met, and further model output still contains a tool call. What is that call's place in this loop?
80.A person must confirm one tool before it runs. What does that confirmation gate decide?
81.A write tool has already returned success. The later final answer describes a different outcome. What is the state of the write?
82.The loop stops after a tool has opened a remote session. What does that stop do to the session?
83.What is working memory for a model call in the loop?
84.What is stored memory?
85.Which of these share the finite context window on one call?
86.The harness saves neither the transcript nor a summary, and a new session starts. What of the old window is in the new call?
87.The harness saves a summary, drops the rest of the transcript, and a later run loads the summary. What is in that later context from the old run?
88.How does a note kept in stored memory enter a later model call?
89.A memory tool saves a note, and its observation is only the word saved. A later session never loads the note. What of the note is in that session?
90.The harness drops older turns so the next call fits the window. A detail in a dropped turn was not saved elsewhere. Where can the model read it?
91.What does planning produce in an agent loop?
92.How does a reactive loop choose its next action?
93.Which pattern is an agent pattern?
94.A plan in the context says to email the customer. What sends that email?
95.A plan is in the context, and the harness has not restricted the tools to that plan. What can the next step call?
96.What makes a setup multi-agent rather than one agent with several tools?
97.One agent can call a read tool and a write tool. What does the second tool add?
98.What does a handoff have to state about the next agent?
99.The parent's reasoning was never copied into the handoff message. What of that reasoning does the child receive?
100.A handoff grants the child some of the parent's tools and leaves the others out. Which of the parent's tools can the child call?
101.From the parent's side, what arrives when the child finishes?
102.The child reaches its own stop and returns a result. What happens to the parent loop?
103.The child returns the word done and omits the figures the parent still needs. Where can the parent read those figures?
104.What weights can the second agent use?
105.Agent A hands work to agent B, and B hands work back to A. What gives that chain a defined end?
106.Code always runs a search, then asks the model to summarize, and the model never chooses the next action. What role is the model playing?
107.What does a search tool supply to the model?
108.Retrieved documents are placed in the prompt, and the model answers in one completion without calling a tool. What kind of step was the retrieval?