In-depth analysis · About 5 min read
AgentCore’s presentation sample is a RAG workflow, not a four-cent slide service
AWS reports a six-slide internal test, with separate model cost, asynchronous delivery and an important OpenSearch network condition.
- Source published
- Published here
- Source checked
An October 14 AWS guide shows how a supervisor agent can turn enterprise documents into a PowerPoint file or a summary. It connects retrieval, text and image generation, document assembly and download through AgentCore and several other services. The interesting result is an end-to-end reference architecture. The reported 70–90 seconds and roughly $0.04 are bounded internal observations, not a public subscription price or a universal service-level guarantee.
Listen to the news
Read this article aloud or keep listening to the news.
Preparing the reader…
Device voice
Speech is handled by your browser or device. AISubHub does not send the text to an audio API. Voice availability and playback with a locked screen vary by device.
Resuming may repeat a short phrase.
At a glance
- The test used one 90-page Amazon 2024 annual-report PDF and produced six slides with two generated images.
- About $0.04 refers to model invocation cost at the time of writing in us-east-1, not the whole cloud deployment.
- Generation exceeds the default 29-second API Gateway timeout; the guide describes asynchronous delivery and five-minute download URLs.
A supervisor routes work instead of writing every component
A Strands-based agent on AgentCore runtime chooses between presentation generation and summarization tools. Requests enter API Gateway and a .NET 8 Lambda Chat API, with conversation state in DynamoDB. The presentation tool invokes a containerized Python Lambda to retrieve context, create outlines and images and assemble PPTX. The summary path queries the knowledge base. S3 holds source and generated documents; Titan Text Embeddings v2 and OpenSearch Serverless support retrieval. The sample uses Claude Sonnet 4.6 for reasoning and content, with separate parsing and image model settings. These are components used by an application, not newly released models or rights granted by a consumer Claude subscription.
Grounding and network isolation require separate checks
The retrieval path parses documents, including tables, charts and images, chunks them, creates embeddings and retrieves relevant context before generation. Grounding can reduce unsupported claims but does not prove every slide is correct. The source also includes a network qualification that should not vanish behind a general claim of VPC isolation: the OpenSearch Serverless collection data plane remains publicly accessible for this architecture because Bedrock Knowledge Bases connects from AWS-managed infrastructure. IAM data-access policies protect the collection, while the dashboard is restricted to VPC access. Our reading is that public network reachability and public data access are different questions; both must be reviewed rather than assuming the entire data path is private.
The reported result belongs to one internal test
AWS reports a 90-page annual-report PDF producing six slides in 70–90 seconds, including retrieval, generation, images and assembly. Two images were generated. A single 1.3 MB PDF was ingested in under two minutes. Around $0.04 in us-east-1 covers model calls at the time of writing, with image generation the largest share; it is not an all-in endpoint, vector storage or operations bill. Our analysis is that an adoption test should keep these scopes separate: ingestion time differs from generation time, generated-image count affects cost, and a short slide deck is not evidence of accurate long-form reporting across an entire document collection.
Long-running work needs an explicit delivery contract
The guide requires an API Gateway Regional REST integration timeout of at least 300 seconds while also explaining that the default is 29 seconds and its Chat API uses asynchronous invocation. Configuration examples still list individual Lambda timeouts. Those statements make request flow a review item, not a reason to copy defaults without checking: clients should use streaming or polling for completion rather than wait on a synchronous REST request. S3 presigned download links expire after five minutes. We would track job status and retry behavior, bind results to the right tenant and avoid exposing bearer download links in public logs. None of those recommendations means we built or tested this queue here.
Security and cleanup are part of the implementation
The sample describes WAF rate controls, Cognito, guardrails, KMS encryption, input validation and per-tenant STS role assumption. Guardrails filter content and sensitive information; they do not replace source verification or tenant authorization. Deployment needs AWS permissions, regional model access, .NET 8, Python 3.12 or later, CDK, an existing VPC and an ARM64-capable container build. Secrets belong in a secrets manager rather than environment configuration intended for resource IDs. Cleanup includes runtime resources, stacks, tenant knowledge bases, buckets and container repositories, with deletion protection handled deliberately. Our conclusion is to evaluate factual fidelity, tenant separation and lifecycle cost alongside slide appearance before considering a production rollout.
Why it matters
This architecture makes document-generation dependencies visible: retrieval, inference, assembly, authorization and delivery each have different failure modes. It is useful as an engineering starting point, especially when a team needs control over source documents, but it should not be compared directly with a flat-price consumer presentation app.
Limits and availability
We read the complete official guide but did not deploy the sample, upload a report, generate slides or verify its timings and bill. Results are AWS’s internal example and prices can vary. No universal accuracy, fully private network path, four-cent total service or new consumer-model entitlement is established.
Sources and evidence
Source published: Oct 14, 2026 · Source checked: Oct 20, 2026
AI-assisted editorial articles checked against linked primary sources. Vendor claims are attributed; inclusion is not endorsement.