Oracle Agentic AI Foundations study plan

Define agentic AI and distinguish it from single-turn prompting, RAG-only chatbots, and classic workflow automation using your own words and one worked example each.

Diagram the core agent loop (perceive, plan, act via tools, observe, reflect) and label where memory, tool calls, and stopping conditions sit.

Compare at least three orchestration patterns (single agent with tools, supervisor with sub-agents, sequential pipeline) and state one failure mode for each.

Write down the vocabulary you cannot yet define precisely (e.g., tool schema, planner, guardrail, handoff) and resolve at least half of it from primary documentation.

Explain how planning, task decomposition, and reflection differ, and identify which of these a given design actually needs.

Map the Oracle AI agent platform's main building blocks (agent definition, tool or function registration, model access, runtime, observability) to the concepts from week 1.

Walk through the platform's documentation to locate where agents are created, configured, and invoked, and note the exact console or API entry points.

Identify which parts of your week-1 diagram the platform provides natively versus which you must build yourself.

Create a minimal working agent on the platform that answers a question using at least one registered tool.

Configure the agent's instructions, model selection, and tool bindings, and record every setting you changed and why.

Test the agent with three inputs: one in scope, one out of scope, and one ambiguous, and log the observed behavior.

Document the end-to-end request path from user input to tool invocation to final response.

Design two tool schemas (one read-only lookup, one action that changes state) with precise names, descriptions, and parameter constraints.

Write tool descriptions that tell the model when NOT to call the tool, and test whether the agent respects that boundary.

Implement error handling for tool failures (timeout, invalid arguments, empty result) and verify the agent recovers or escalates cleanly.

Connect the agent to at least one external system or API and confirm authentication and payload shaping work end to end.

Explain the difference between retrieval-augmented generation and agentic retrieval, and state when each is appropriate.

Chunk, embed, and index a small document set, then measure retrieval quality on at least ten hand-written queries.

Give the agent a retrieval tool and tune top-k and similarity thresholds based on the measured results.

Identify and mitigate one grounding failure (stale chunk, wrong document, or unsupported claim) and record the before-and-after behavior.

Enumerate the trust boundaries in your agent (user input, tool output, retrieved content, model output) and name the risk at each boundary.

Implement at least two guardrails: input validation or prompt-injection screening, and output filtering or action confirmation for state-changing tools.

Apply least-privilege credentials to each tool and verify the agent fails safely when a permission is missing.

Write a short responsible-AI note covering data handling, human-in-the-loop points, and known limitations of your agent.

Instrument the agent so each run records inputs, tool calls, latency, token usage, and outcome.

Build a small evaluation set of at least fifteen cases with expected outcomes and run it as a regression suite.

Define three operational alerts or thresholds (error rate, latency, cost per run) and state the action each triggers.

Describe a versioning and rollback approach for agent instructions, tool schemas, and models.

Re-review every domain against the exam weighting and list your three weakest areas with a concrete fix for each.

Take a timed practice set or write and answer fifty recall questions spanning all six domains.

Re-explain the full lifecycle of your demo agent from design through monitoring in under five minutes without notes.

Close remaining gaps in the vocabulary list from week 1 and confirm every term is now defined precisely.