Understanding context engineering, RAG and least privilege is not the same as understanding them. The direct method is to run the same task once without, then once with, each design. These are pre-scripted teaching simulations: they call no real model, collect no input and execute no outside action.
00 / METHODChange only one condition at a time
Run the “without” condition first, then the “with” condition on the same sentence. Do not watch only the final answer: the difference occurs earlier, in what the agent sees, its first decision, its tool call, its return to the result and its ability to stop before danger.
01 / COMPARISONContext ablation: how much does the model actually see?
Task: organize a new case. Without structured context, the system sees one request but no parties, issue, materials, deadline or verified facts. It fills gaps with a generic civil-case checklist and cannot verify the result. With a materials list, confirmed facts, open checks, deadline and prohibitions, it identifies a termination-date and service-evidence gap before drafting, creates a timeline separated into confirmed, stated and unverified facts, and excludes an unnecessary identity scan. The result is a next action that fits this case: verify service date and obtain April–June payroll records.
02 / COMPARISONRAG comparison: answer first or find the source first?
Task: ask when an employee must submit reimbursement. Without retrieval, the model guesses an ordinary month-end rule although the company policy is absent; the fluent answer has no version, clause or exception. With retrieval, it finds Expense Reimbursement Policy V3.2, rejects retired V2.8, locates Article 12’s 30 calendar days and checks the travel exception. The answer cites version, clause and effective date, then says the calculation changes if this is travel expense. RAG adds traceability and a way to challenge or update an answer.
03 / COMPARISONTools and guardrails: being able to act is not a reason to act now
Task: send organized materials to a client. Without permissions and verification, organizing and sending form one action: the system assumes the most recent contact and file are correct, sends immediately, and leaves only a receipt that cannot prove correct recipient or attachment. With least privilege, read, draft and send are separate. The system creates a preview, detects an identity page in an attachment, removes it, and stops at a reversible confirmation point. It states that nothing has been sent and names the recipient and redacted attachments for confirmation.
04 / OBSERVEDo not ask which paragraph sounds better. Ask which trace is more reliable.
- Context moves a model from generic advice to the current task, but stale context can consistently mislead it.
- Retrieval makes an answer return to a version, clause and source; when nothing matches, a reliable system says it does not know.
- Guardrails move irreversible action behind preview and confirmation. Sometimes an agent’s refusal to act is the correct result.
The three comparisons make the same point: an agent’s capability comes not only from the model but also from the working environment around it.
Sources
- bojieli/ai-agent-book.
- Context engineering, memory and RAG, and tools and MCP.
- Cases, traces and answers are pre-scripted teaching simulations. They are not model evaluation and perform no external action.
READER COMMENTS
Leave the thought this article gave you.
No comments yet. You can leave the first one.