The DBA’s Guide to Oracle Agentic AI Foundations (Reasoning patterns, memory and guardrails)

Agentic AI Foundations Series · Part 2 of 6

Part 2: How Agents Think – Reasoning Patterns, Memory, and Guardrails

1Z0-1157-26 exam prep for Oracle DBAs

Agents follow a few well-known thinking styles. Learn the name, the one-line idea, and the keyword that gives it away in a question.

Five thinking styles, one DBA story

The problem: "Users say the month-end report is slow."

• Chain-of-Thought (CoT) – You think aloud before answering: "Slow report → maybe bad plan → maybe stale stats." No tools, just reasoning step by step.
• ReAct (Reason + Act) – You think, run something, look, think again. Thought: check the plan. Action: run DBMS_XPLAN. Observation: full table scan. Thought: check stats… This is the most tested pattern.
• Plan-and-Execute – You write the full plan first (1. AWR, 2. SQL plan, 3. stats, 4. fix), then work through it. Good for long tasks.
• Reflection – After writing your RCA, you review it yourself: "Did I prove the root cause or just guess?" The agent critiques and improves its own output.
• Routing – A triage step sends the issue to the right specialist: performance, storage, or EBS.

Exam hook: Thought-Action-Observation = ReAct. "Plan first" = Plan-and-Execute. "Critic" = Reflection.

When one agent isn't enough

Think of a war room during a P1 outage.

• Supervisor pattern – the incident manager assigns work to the DBA, network, and app teams, then combines their findings.
• Handoff – L1 support passes the ticket (with full history) to L2. L1 is now out of it.
• Agent as a tool – the incident manager asks the network team one question, gets the answer, and stays in charge.

Multi-agent gives specialization and parallel work, but costs more tokens, time, and debugging effort.

Memory: what the agent remembers

• Short-term – tonight's conversation, held in the context window. Gone when the session ends.
• Long-term – saved in a database or vector store and recalled by meaning.
• Sub-types you may see: episodic (past events), semantic (facts), procedural (how to do things).

Guardrails: change control for agents

You would never let a junior DBA run anything in production without controls. Same for agents:

• Input guardrails – block prompt injection, jailbreaks, off-topic requests.
• Output guardrails – catch PII leaks, toxic text, wrong formats, hallucinations.
• Least privilege – a read-only DB user for a reporting agent.
• Human approval for anything irreversible: delete, pay, send.
• Limits – max loop iterations, timeouts, token budgets.
• Tracing – log every step, like an audit trail.

Remember for the exam

• ReAct = Thought → Action → Observation.
• Handoff gives control away; agent-as-tool keeps control.
• Short-term = context window; long-term = vector store/DB.
• Guardrails sit on both input and output.

Self-check

1. An agent that drafts an answer, then critiques and rewrites it, uses…? (Reflection)
2. A manager agent delegating to specialist agents is the…? (Supervisor pattern)
3. Where does short-term memory live? (The context window)

Next in the series: Part 3