What happened
AWS’s announcement adds two modes to Amazon Bedrock Managed Knowledge Base’s AgenticRetrieveStream API. The service already supported agentic retrieval, in which a foundation model breaks a question into sub-queries, searches knowledge bases, assesses whether the evidence is sufficient, and can refine its searches. The new maxAgentIteration range is 0–5; AWS says existing requests using values 2–5 retain their current behavior.
At maxAgentIteration=1, the service retrieves once and generates one answer grounded in the retrieved passages. There are no planning rounds or query rewrites. AWS describes this as single-shot RAG and suggests it for direct, well-scoped questions. The setting can use multiple knowledge bases, as can the other modes. AWS says one foundation-model call generates the answer and that avoiding agentic planning can reduce latency and model usage compared with agentic retrieval.
At maxAgentIteration=0, the API performs one retrieval and returns passages without generating an answer or invoking a foundation model. This resembles a direct Retrieve API call, though AWS says the streamed response format differs: passages arrive in a result event without score or location fields. Setting generateResponse=true with zero iterations is invalid and returns a 4xx error.
For values 2–5, the service remains in agentic mode. AWS explains that one iteration equals one foundation-model call and the final iteration is reserved for answer generation. A value of *n* therefore permits up to *n* minus 1 planning rounds. Each planning round combines a query rewrite with retrieval; an initial speculative retrieval that runs ahead of planning does not count toward the limit. AWS illustrates the choices with a fictional banking assistant: a checklist lookup uses 1, a cross-policy comparison uses 4, and a source-text review uses 0. The example is not a customer deployment or measured result. The application selects the mode; the setting does not provide automatic routing.
What the archive adds
New Runtime’s earlier coverage includes a November 6, 2026 note on Web Answer in AgentCore, a distinct AWS-managed tool for web search. That is adjacent AWS agent infrastructure, not the same product or capability. This announcement is a new event: it adds retrieval modes to Managed Knowledge Base’s AgenticRetrieveStream, rather than extending Web Answer. Read the earlier note.
Trend signal
This does not yet fit a pattern New Runtime tracks. The change gives application builders explicit controls over retrieval effort, but AWS does not describe a general interoperability layer, runtime security boundary, or completed-task pricing model. Its practical significance is local to choosing among retrieval behaviors inside one Bedrock API.
Why now
AWS frames the change around a common application design problem: the same assistant may receive both direct questions and queries requiring evidence across multiple documents. The new settings let an application use different retrieval effort without maintaining separate API paths for simple and complex questions.
Why it matters
Retrieval systems often make a tradeoff between quick answers and broader evidence gathering. With these settings, a team can keep a direct lookup from entering a planning loop, while allowing comparisons and exploratory questions to search in stages. A retrieval-only path also lets an application present source text for human review without paying for answer generation through a foundation-model call.
That control is useful only if the application can classify or route requests appropriately. AWS explicitly leaves mode selection to the application, so teams still need to decide which requests are straightforward and which need multi-step retrieval. The announcement does not establish that single-shot mode is cheaper overall or equally accurate; it says it can reduce latency and model usage, and advises evaluation against representative questions. Builders should compare modes using their own corpora, including questions whose answers span documents and questions the corpus cannot answer. The goal is not to maximize iterations, but to find the lowest setting that meets each workflow’s quality requirements.
Limits and uncertainty
AWS’s banking assistant is illustrative, not a reported customer deployment, and the announcement provides no measured latency, cost savings, or accuracy results. The benefit of single-shot retrieval depends on whether one search finds adequate evidence; the benefit of additional planning depends on the question and the quality of the indexed material. AWS recommends evaluating representative questions, but the announcement does not quantify a quality threshold or specify a price for each mode. The API’s behavior is described by AWS and has not been independently tested here.
Skill to build now
Build a small routing evaluation set this week: include direct lookups, questions requiring comparisons across documents, and questions with no answer in the corpus. Run each through iteration limits 0, 1, and an appropriate agentic setting; record latency, request cost, answer correctness, and whether cited passages support the answer. Use those results to set application-side routing rules rather than assuming a higher iteration count is always better.
Reader takeaway
AgenticRetrieveStream can now serve three retrieval needs through one API: passages only, one retrieval plus a generated answer, or multi-step agentic search. The useful change is control over effort per request—not automatic routing. Measure quality and latency on your own documents before changing defaults.