NVIDIA NCP-AAI study plan

Map the anatomy of an agentic loop (perceive, plan, act, observe, reflect) and label where tool calls, memory reads, and termination checks occur

Compare single-agent and multi-agent topologies (supervisor, hierarchical, peer-to-peer handoff, blackboard) on coordination cost, failure isolation, and debuggability

Write explicit role, responsibility, and handoff contracts for a three-agent pipeline so each agent's input and output schema is unambiguous

Identify the top five failure modes of agent handoffs (context loss, circular delegation, role drift, unbounded retries, ambiguous termination) and propose one mitigation per mode

Design tool and function interfaces with typed parameters, idempotency keys, and structured error returns so agents can recover from failed calls

Specify memory layers (working context, episodic store, long-term retrieval) and decide what each agent may read and write

Define orchestration control flow including retry budgets, timeouts, escalation paths, and human-in-the-loop checkpoints

Draft evaluation criteria for multi-agent traces: task success, handoff accuracy, redundant-step rate, and cost per resolved task

Characterize how latency, token cost, and context-window pressure scale as agent count and step depth grow, and identify the dominant bottleneck in each regime

Select parallelism and concurrency patterns (fan-out/fan-in, worker pools, async queues) that fit the workload's dependency graph

Design state management for long-running agent sessions, including checkpointing, resumability, and idempotent replay after failure

Plan horizontal scaling of tool and model endpoints with rate-limit handling, backoff, and request coalescing

Containerize an agent service and expose a health endpoint, structured logs, and trace identifiers that propagate across agent hops

Instrument the pipeline with distributed tracing and per-step metrics so a single task can be reconstructed end to end

Configure autoscaling rules and load-test the service to find the concurrency level where p95 latency degrades

Write a runbook for common production incidents: stuck loops, tool-endpoint outages, and runaway token spend

Enumerate the ethical risk surface of agentic systems (harmful tool use, data leakage across agents, bias in routing, opacity of decisions) and rank each by likelihood and impact

Define guardrail layers: input validation, tool-permission scoping, output filtering, and escalation to a human reviewer

Specify an audit-log schema that captures decision rationale, tool invocations, and approver identity for every consequential action

Draft a governance policy covering agent ownership, change approval, and periodic review cadence

Build adversarial test cases that probe prompt injection, tool-argument tampering, privilege escalation, and cross-agent data leakage

Implement least-privilege tool scoping and verify it blocks a deliberately over-broad request

Define human-in-the-loop gates for irreversible or high-impact actions and test that the gate cannot be bypassed

Establish a red-team review cadence with pass/fail criteria tied to the risk register

Assemble an end-to-end evaluation harness that scores task success, tool-call correctness, handoff quality, latency, and cost on a fixed task set

Run regression comparisons between two prompt or model configurations and interpret whether differences exceed noise

Tune retrieval and context assembly to reduce redundant steps without lowering task success

Document deployment stages (dev, staging, canary, production) with promotion criteria and rollback triggers

Consolidate weak areas by re-scoring practice questions per domain and re-studying any domain below a self-set threshold

Rehearse explaining architecture trade-offs, scaling bottlenecks, and governance controls aloud in exam-style scenarios

Verify recall of key definitions: agent loop, handoff contract, idempotency, least privilege, audit trail, human-in-the-loop

Finalize a personal one-page reference sheet of decision rules for design, scaling, and ethics questions