AI agents basics
An AI research scientist interview on the agent patterns and measures a study treats as known.
These prompts are original practice. They are not questions from a named employer or a live hiring loop.
1.What does an agent policy select from a state?
2.For a tool-using language-model agent, what plays the role of the state?
3.Which of these is an action for a tool-using language-model agent?
4.What does a stochastic agent policy assign over actions in a state?
5.What does a deterministic agent policy name for one state?
6.A tool call names a tool and also supplies arguments. What do the arguments belong to?
7.What further tool result does the final-answer action request?
8.Two episodes use the same weights and have different transcripts. What differs for the policy?
9.What does the ReAct pattern interleave?
10.In ReAct, where is the reasoning step placed relative to the action it refers to?
11.In a ReAct loop, when is a tool result first available to a later reasoning step?
12.What does using the ReAct pattern establish about the final answer?
13.A transcript has actions and observations, and no reasoning step before the actions. What pattern is that?
14.How does a block of reasoning that never calls a tool differ from ReAct?
15.What order does plan-and-execute use?
16.Which description fits a reactive policy?
17.A plan is written and none of its steps have been carried out. What is that plan?
18.What is true of observations that arrive during open-loop execution of a plan?
19.Under closed-loop execution, what can the latest observation affect?
20.What is replanning?
21.What does reflection add about an attempt that already happened?
22.A critique is written and no training step is run. What is the critique?
23.When does reflection come together with a change in the weights?
24.A critique is generated and then withheld from the next attempt. What does that attempt condition on?
25.A critique is present in the next attempt's context. What may still be true of the critique?
26.Several reflection rounds run and no training step runs. What has accumulated?
27.What is a trajectory?
28.The same task is started a second time. What is the trajectory of the second run?
29.What does a final-answer metric leave out of the score?
30.What does a process metric score?
31.The final answer is right, and the path would fail a process check. How can a final-answer metric treat the episode?
32.The path meets a process metric and the final answer is wrong. What can the two metrics do?
33.Two trajectories end with the same answer. What can a process metric still separate?
34.What does trajectory length count?
35.How is success on a task suite expressed?
36.What does that suite success rate leave unidentified?
37.A task succeeds and its trajectory contains a call to one tool. What does that call show about necessity?
38.In an unweighted success rate, how does each task contribute?
39.A final answer is correct and no tool was called. How does a final-answer success rate count that task?
40.A recorded label judges only whether the finished result is acceptable. No step carries its own label. What kind of label is that?
41.Judgments are written on intermediate steps, whether or not the finished result has a separate judgment. What are those step judgments?
42.Why is an outcome label a different label from a process label?
43.An outcome label says the end result failed. Which step does that label identify as the faulty one?
44.A step is labeled acceptable and the end result is labeled unacceptable. What does that combination show?
45.A training set carries an outcome label for each finished answer and no other labels. What is missing unless it was collected on its own?
46.What can a tool-use learning step change, when a training signal about tool use is applied?
47.Tools appear only as descriptions in the prompt, and no training step runs. What changes?
48.What is a few-shot example of a tool call that is written in the prompt?
49.Stored tool-call traces are used in a gradient step. What kind of step is that?
50.The prompt describes a tool, and the agent never calls it. What is the description?
51.Fine-tuning on tool-call traces is combined with rewritten tool descriptions. What has changed?
52.How does a tool description in the prompt enter the episode?
53.The weights stay frozen and the tool description in the prompt is replaced. What is true of the episode?
54.A skill is present only as instructions written in the prompt. What is that skill?
55.Which written pieces of a tool description are part of the episode's condition?
56.For a difference in actions to be attributed to the weights alone, what else has to have stayed fixed?
57.What is multi-agent debate?
58.Debating agents settle on one claim. What is that settlement?
59.What does a majority vote over debate messages check?
60.Every debater accepts a claim, and no observation in the transcripts supports that claim. What is the acceptance?
61.Debate transcripts are not merged. What state does each debating policy condition on?
62.A handoff does not say that any state is shared. What does the receiving agent start from?
63.Which of the sender's records reaches the receiver only if the handoff payload includes it?
64.What sets the receiving agent's starting state after a handoff?
65.After a handoff, the receiver calls tools. When do those tool results enter the sender's state?
66.What is in the action space of a tool-using agent that is allowed to finish?
67.One new tool is made available. The weights, the task, and the other tools stay as they were. What changed?
68.One tool is withdrawn and nothing else is changed. What is different about the actions the agent can take?
69.Two calls use the same tool name and different arguments. Are they the same action?
70.Two setups allow a stop action and give the agent different tools. Do they share one action space?
71.What does the observation function determine?
72.A tool returns an error, and the agent is shown a blank field instead. What changed?
73.The agent is shown a summary of a tool result rather than the result itself. What differs?
74.An observation omits part of the environment state. What can the policy read from the omitted part?
75.The model writes a paragraph, and the loop files that paragraph as the tool's return. What is the paragraph?
76.A trajectory would have continued, and a step limit stops it. What did the limit do to the later steps?
77.A run reaches the step limit before any final answer. How does a success rate count that run?
78.The same policy is scored under two different step limits. What do the two success rates represent?
79.Relative to the model parameters, where is a step limit defined?
80.A non-idempotent tool is called again with the same arguments. Why is that retry not the same action as the first call?
81.What has to be said about retries before the actions available in the episode are fully specified?
82.How does repeating an idempotent read differ from retrying a non-idempotent tool?
83.Allowing or forbidding a retry is which kind of specification?
84.What does grounding require an answer to be tied to?
85.An answer is right because of the parameters, and no observation in the trajectory contains it. Is that answer grounded in the trajectory?
86.The tool returned an error. The answer quotes a result the tool never returned. Is the quotation grounded?
87.Can an answer be grounded and still be wrong?
88.A memory write remains after the episode that wrote it. What can a later episode do with that write?
89.The window is cleared at the end of each episode. What does the next episode fail to inherit from that window?
90.How do persistent memory and a window cleared every episode differ for the next episode?
91.Must memory that persists across episodes be a weight update?
92.Two episodes use the same prompt and the same weights. The window is cleared between them, and nothing else is persisted. How do their starting transcripts compare?
93.What does an ablation do, as a comparison?
94.When is removing a tool an informative ablation of that tool?
95.An ablation removes one tool under one fixed policy and one fixed task. What does that comparison not establish?
96.A condition with one factor removed is reported alone, with no paired condition. Why is that not an ablation?
97.Which roles define a hierarchical agent setup?
98.Do the callee's intermediate tool results automatically appear in the caller's context?
99.The callee returns a short result and does not include its tool trace. What of that trace does the caller then have?
100.From the caller's side, what is the invocation of the callee?
101.What does a reward model assign to an output?
102.When is a reward model also the agent policy?
103.A reward model scores an output, and no training step uses the score. What happens to the weights?
104.What can a training step that uses a reward score do, beyond computing the score?
105.What does raising the temperature do to the distribution the next action is drawn from?
106.What does sampling do with the policy's distribution over the next action?
107.While greedy decoding is writing an action, what does it select at each step?
108.A trajectory is given without the temperature and without the sampling rule. What about the policy is under-specified?