Agent Design and Cognition
Free NVIDIA-Certified Professional: Agentic AI practice — 6 questions on Agent Design and Cognition, with explanations. No sign-up. Full 12-question mixed test →
An agent must solve a puzzle where an early move can be wrong and only becomes apparent several steps later, requiring exploration of multiple candidate move sequences and the ability to abandon bad branches mid-solution. Which reasoning paradigm BEST fits this requirement?
Tree-of-Thought explicitly maintains multiple partial solution branches and allows evaluating and discarding weak branches, which matches the need to backtrack after a bad early move.
A finance team needs an agent to generate a quarterly report by pulling data from four known APIs in a fixed order, aggregating results, and producing a PDF. The steps and dependencies are fully known in advance and do not change based on intermediate results, and minimizing per-step LLM token cost is a priority. Which control pattern is BEST?
Since the steps and dependencies are fully known and unlikely to change, planning once and executing deterministically avoids repeated LLM reasoning calls, directly satisfying the cost-minimization requirement.
You are designing a tool-calling schema for an agent that invokes a 'create_refund' function. To minimize the chance the model invokes the function with an invalid or missing amount field, which schema design choice is MOST effective?
Enforcing a required, typed field at the JSON schema level allows the calling framework to structurally validate and reject malformed calls before execution, independent of prompt phrasing.
A company wants an assistant that independently reasons about legal compliance, technical troubleshooting, and billing disputes, each requiring distinct specialized prompting, tool access, and safety guardrails. Users submit tickets covering all three domains simultaneously. Which architecture is BEST?
Distinct specialized personas, tools, and guardrails per domain, combined with the need for a coordinated final answer, argue for an orchestrator delegating to domain-specialist sub-agents.
An agent architecture adds a self-critique step where the model reviews and revises its own draft answer before returning it. Which statement about this pattern is MOST accurate for exam purposes?
Self-critique adds an extra generation pass, raising latency and cost, and is effective at catching reasoning/format errors, but it has no access to ground truth, so it cannot reliably fix factual hallucinations without external grounding.
An agent is being designed to execute stock trades exceeding $50,000 based on market signal analysis. Regulatory requirements mandate human review before any irreversible high-value action. Which control design is BEST?
A hard, rule-based deterministic checkpoint reliably enforces the regulatory mandate for human review before irreversible high-value actions, with the agent constrained to only proposing the trade.
Ready for the real thing?
The full course: two full-length practice tests, video lessons for every exam domain, hands-on labs and detailed explanations.
$129.99 $34.99 with code FREETEST33 — valid through September 7.
Get my $34.99 deal →