Architecture AI Updates: October 25, 2026

1. RAG vs. Agentic RAG: A Fixed Pipeline vs. a Retrieval Loop With a Sufficiency Check

ByteByteGo. ByteByteGo compared standard RAG, which retrieves context once from a single knowledge base and passes it to the LLM, with agentic RAG, where an agent picks tools (documents, APIs, databases, web search), checks whether the evidence is sufficient, and adjusts its query or tool before generating an answer. The main differences it lists are single versus repeated retrieval, fixed versus adaptive control flow, one versus many source types, and the lack of any built-in quality check in standard RAG. A reader comment points out a failure mode: when the knowledge base cannot answer a question, the sufficiency check may never pass. The commenter recommends a search budget, logging why each run stopped, and returning “not found” once the budget runs out. Source

2. Advisory Skills for Agent Behavior, Enforced IAM for Agent Permissions

InfoQ. InfoQ’s coverage of Google’s Data Agent Kit GA describes two layers of guardrails for coding agents working against data platforms. Skills such as accidental-data-loss-prevention tell the agent to stop and get explicit user consent before irreversible operations like DROP TABLE or unfiltered DELETEs, but nothing enforces them. Read-only isolation is enforced in IAM instead, through read-only service accounts or a deny policy on MCP tool calls where tool.isReadOnly is false. The article lists limitations: blocked tools still appear in tools/list, and tool results carry no idempotency key or commit status that would show whether a timed-out write actually committed. Source