Agents Interview Questions

Summary

Interleave reasoning traces with tool actions in a loop until a final answer.

Longer Answer

Interleave reasoning traces with tool actions in a loop until a final answer. Good for short-horizon exploratory tasks.

Key Points

• Interleave reasoning traces with tool actions in a loop until a final answer
• Good for short-horizon exploratory tasks
• Core agent pattern
• Use the answer to connect the concept back to a realistic production or interview scenario for 'Agents leaf: What is the ReAct pattern?'

Comparison / Tradeoffs

Summary

When the workflow is known, latency/cost budgets are tight, or side effects need strict control-use single-shot LLM or deterministic code.

Longer Answer

When the workflow is known, latency/cost budgets are tight, or side effects need strict control-use single-shot LLM or deterministic code. Architecture judgment.

Key Points

• When the workflow is known, latency/cost budgets are tight, or side effects need strict control-use single-shot LLM or deterministic code
• Architecture judgment
• Relevant areas: Ai, Agents, Architecture
• Use the answer to connect the concept back to a realistic production or interview scenario for 'Agents leaf: When should you avoid agents?'

Comparison / Tradeoffs

Summary

Max steps, token budgets, repeat-detection, timeouts, and forced finalization.

Longer Answer

Max steps, token budgets, repeat-detection, timeouts, and forced finalization. Log every iteration for debugging.

Key Points

• Max steps, token budgets, repeat-detection, timeouts, and forced finalization
• Log every iteration for debugging
• Control loop hygiene
• Use the answer to connect the concept back to a realistic production or interview scenario for 'Agents leaf: How do you prevent infinite tool loops?'

Comparison / Tradeoffs

Summary

Model emits name+args matching a schema; runtime validates and executes.

Longer Answer

Model emits name+args matching a schema; runtime validates and executes. Prefer JSON schema enforcement over free-form text.

Key Points

• Model emits name+args matching a schema; runtime validates and executes
• Prefer JSON schema enforcement over free-form text
• Tool calling basics
• Use the answer to connect the concept back to a realistic production or interview scenario for 'Agents leaf: What is structured function calling?'

Comparison / Tradeoffs

Summary

Bound the loop in code: a maximum number of steps per task type, a token or cost budget, detection of repeated identical tool calls, and a wall-clock limit, with a graceful ending (summary and handoff) when a limit is hit. Also make tool results informative so the model has a better next action than retrying.

Longer Answer

The naive loop stops only when the model stops asking for tools, which a confused model may never do. Empty results like [] invite retries; "no orders match; order numbers have 8 digits" guides the next step. Track steps and cost per conversation and alarm on outliers.

Key Points

• Step, cost and time budgets in code.
• Detect repeated calls.
• Informative tool results.
• Graceful ending and alerts.

Deep dive: The Agent Loop and the Runaway Bill

Summary

Classify tools by risk (read-only, reversible, destructive, external), enforce approval in the host's code for destructive and external actions, show approvers the exact arguments and blast radius, approve plans before execution for multi-step changes, and prefer dry runs, per-call limits and reversible operations.

Longer Answer

Approval enforced in a prompt can be talked around; approval enforced in code cannot. Too many prompts cause rubber-stamping, so auto-approve read-only and small reversible actions. Log approver, arguments and outcome.

Key Points

• Risk classes per tool.
• Gates in code.
• Show exact action and blast radius.
• Dry runs, limits, reversibility.

Deep dive: Human Approval for Risky Actions

Summary

On outcomes, not exact paths: run tasks in resettable sandbox environments, check the final state in code (record created, ticket closed), check constraints (no forbidden actions, approvals obtained), measure efficiency against budgets, and run each task several times to report success rates.

Longer Answer

Agents reach goals through different valid paths, so matching a golden tool sequence fails good runs and passes runs that look right but produce wrong results. Grade the final message with a rubric where it matters.

Key Points

• Grade final state and constraints.
• Efficiency against budgets.
• Repeated runs for success rates.
• Sandboxed, resettable environments.

Deep dive: Evaluating Agents

Summary

Model invents tool names/args.

Longer Answer

Model invents tool names/args. Mitigate with strict schemas, constrained decoding, and runtime rejection of unknown tools.

Key Points

• Model invents tool names/args
• Mitigate with strict schemas, constrained decoding, and runtime rejection of unknown tools
• Common failure mode
• Use the answer to connect the concept back to a realistic production or interview scenario for 'Agents leaf: What is tool hallucination?'

Comparison / Tradeoffs

Summary

In a workflow, your code defines the sequence of LLM calls and tools (prompt chaining, routing, parallelisation); in an agent, the model decides the next step dynamically in a loop. Workflows are more predictable and cheaper; agents handle open-ended tasks where the steps can't be known in advance.

Longer Answer

Many production systems are workflows with one agentic step. Prefer the simplest design that works, and add autonomy only where it measurably helps.

Key Points

• Workflow: code decides the steps.
• Agent: model decides the steps.
• Workflows are predictable and cheaper.
• Use the simplest design that works.