Key takeaways

  • AWS published the architecture on October 14, 2026. It routes requests through one AgentCore supervisor to PowerPoint-generation and summarization tools, with Bedrock Knowledge Bases and OpenSearch Serverless supplying retrieved context.
  • AWS reports that one 1.3 MB, 90-page Amazon annual-report PDF produced a six-slide deck in 70–90 seconds. That is a single disclosed internal test, not a latency distribution or a quality benchmark.
  • The post names Claude Sonnet 4.6, Titan Text Embeddings v2 and Stable Image Inpaint. The public repository is available under MIT-0, but its pinned README still shows older model examples, so deployers must reconcile code, documentation and regional model availability.
  • The approximately $0.04 model-invocation claim needs remeasurement: AWS’s current pricing page lists Stable Image Inpaint at $0.07 per generation, while the sample result says it generated two images. Infrastructure, storage, retrieval, network and observability costs are also outside that headline.
  • Treat this as a useful starting architecture, not a ready procurement result. Run representative documents, citation checks, tenant-isolation tests and a complete cost ledger before exposing generated presentations to clients.
01

AWS published the AgentCore presentation sample on October 14

AWS published a reference architecture on October 14, 2026 for turning enterprise documents into PowerPoint files. A supervisor agent hosted on Amazon Bedrock AgentCore Runtime chooses between generate_pptx and generate_summary tools. The generation path retrieves source material, asks models to produce an outline and content, creates images and assembles a .pptx file that is returned through a short-lived Amazon S3 presigned URL.

This is not a new foundation model or a fully managed presentation product. It is an AWS-authored technical sample that customers deploy into their own accounts. The named stack includes Claude Sonnet 4.6 for generation, Titan Text Embeddings v2 for retrieval, Stable Image Inpaint for images, Bedrock Knowledge Bases, OpenSearch Serverless, Lambda, API Gateway, DynamoDB, S3 and AgentCore Runtime.

02

The 70–90-second result covers one disclosed document

AWS says its internal test ingested a 1.3 MB, 90-page Amazon 2024 Annual Report PDF in under two minutes and generated a six-slide presentation in 70–90 seconds. The output included title, bullet, image-and-text, image-only and takeaway layouts, two generated images and speaker notes.

The post does not report repetitions, percentile latency, concurrent users, cold starts, retrieval hit quality, citation correctness, numerical accuracy, slide-edit effort or failure rate. It also uses an Amazon-authored document inside an Amazon-authored test. The result demonstrates that the assembled path completed once under the disclosed conditions; it does not predict performance or quality for scanned files, complex tables, mixed languages, private templates or many simultaneous tenants.

03

Grounding is a pipeline property, not a guarantee

Documents are stored in S3, parsed and chunked, embedded with Titan Text Embeddings v2 and indexed in OpenSearch Serverless. At request time, Bedrock Knowledge Bases retrieves passages and supplies them to Claude. The generated file can contain citations and source-derived summaries, but retrieval-augmented generation only provides context; it does not prove that every claim, number, chart or image caption is supported.

A practical evaluation should extract every factual statement and number from each generated slide, map it to an exact source page or table and score support, contradiction, omission and unverifiable synthesis. Presentation quality needs a separate review for hierarchy, legibility, template compliance and whether a generated visual could be mistaken for source evidence.

04

The public code exists, but the prose and artifact are not one frozen release

AWS links a public aws-samples repository under the MIT-0 license. The GitHub API returned commit 9c51d2c881e9edbe81adbc84fedfa89cfa1b392b as the latest default-branch commit when checked, and that commit can be pinned for review. The README lists .NET 8, Python, AWS CLI, Node.js, CDK, Docker-compatible image building, S3 and an API Gateway timeout quota increase as prerequisites.

The October blog instructs readers to configure Claude Sonnet 4.6 and Stable Image Inpaint, while the pinned README example still names Claude 3.5 Sonnet and Titan Image Generator v2 and omits some parameters shown in the post. That does not make the sample unusable, but it means the article, repository, model access and region matrix must be reconciled before deployment. Pin the exact commit and container digests; do not treat the moving default branch as a release artifact.

05

The $0.04 model-cost claim needs a fresh bill

AWS reports approximately $0.04 in model invocation costs for its example in us-east-1 and says image generation accounts for most of that amount. The same result describes two images generated with Stable Image Inpaint. On the pricing page retrieved October 15, AWS lists Stable Image Inpaint at $0.07 per generation in N. Virginia, Ohio and Oregon. Two listed image generations alone would be $0.14 before text generation.

AccessAllGPT cannot resolve whether the internal test used a different model, an earlier price, an unreported discount, cached artifacts or a mistaken estimate. The current pricing page also separately meters managed knowledge-base storage and retrieval, and the full system adds OpenSearch Serverless, Lambda, API Gateway, AgentCore, DynamoDB, S3, KMS, WAF, Cognito, logging and data transfer. Buyers should measure an itemized account bill per accepted deck rather than reuse $0.04.

06

The security design is substantial but not self-validating

The sample describes VPC isolation, KMS encryption, IAM least privilege, WAF rules and rate limits, Cognito threat protection, S3 public-access blocks and access logs, DynamoDB point-in-time recovery and deletion protection, X-Ray tracing, input validation and per-tenant STS role assumption. AgentCore Runtime separately documents HTTP, MCP, A2A and AG-UI service contracts with SigV4 or OAuth options, depending on protocol.

AWS also notes an important boundary: the OpenSearch Serverless collection data plane must remain publicly accessible for Bedrock Knowledge Bases, with IAM data policies providing protection while the dashboard stays VPC-only. Production review must test document and index isolation, forged tenant identifiers, presigned-URL leakage, prompt injection from retrieved files, citation spoofing, malicious office content, log redaction, role-assumption scope and cleanup after failed jobs. A diagrammed control is not evidence that its policy is correct.

07

What teams should do next

Trial the pinned sample with a small, representative corpus: clean PDFs, scans, tables, charts, contradictory revisions and documents containing adversarial instructions. Freeze the region, models, inference profiles, repository commit, prompts and templates. Run repeated warm and cold tests, then record ingestion and generation latency distributions, retrieval support, citation accuracy, slide-edit minutes, failed jobs and the complete AWS bill.

Constrain the pilot to non-sensitive documents until tenant and data-lifecycle tests pass. Require human approval before a deck leaves the organization. Wait if current model identifiers or regional access cannot be reconciled with the code. Reject client-facing automation when the system cannot trace every material claim to a source, isolate tenants, remove expired artifacts or produce a cost and failure record per deck.

08

Copy-ready document-to-deck trial record

Complete one record for a pinned architecture, document class and presentation template. Treat AWS’s internal result as a hypothesis, not your baseline.

Entries stay in this browser tab and are not submitted to AccessAllGPT. Blank responses are copied as [Unresolved].

Repository commit, container digests, CDK version, region, model and inference-profile IDs, prompt and template hashes.

Document classes, sizes, languages, scans, tables, charts, revision rules, sensitivity, tenant and prohibited content.

Claims and numbers checked, exact source locations, supported, contradicted, omitted and unverifiable counts, citation precision and reviewer.

Layout and brand checks, legibility, accessibility, image provenance, edit minutes, accepted deck rate and rejection reasons.

Ingestion and generation p50/p95/p99, warm and cold runs, concurrency, timeouts, retries, failed jobs and recovery behavior.

Identity path, role scopes, index and object isolation, injection tests, URL expiry, office-file scanning, log redaction and deletion evidence.

Model calls by model, image generations, AgentCore, retrieval, OpenSearch, compute, storage, network, security, observability and review labor per accepted deck.

Trial, constrain, ship, wait or reject; approved data and audience; spend and volume ceiling; human approval; rollback owner; expiry and retest triggers.

Attach source documents, generated decks, claim-to-source records, traces, policy exports and itemized bills. Do not replace measured cost with the post’s $0.04 estimate.

Primary sources

Browse the publication-wide evidence index →

  1. Generate presentations with Amazon Bedrock AgentCoreAWS Machine Learning Blog · Reviewed: October 14, 2026 publication date; solution overview; AgentCore supervisor and tool routing; RAG flow; model identifiers; security controls; deployment prerequisites; internal test setup and results; cost statement; asynchronous API pattern; cleanup and security considerations · Retrieved · Supports: AWS published a deployable reference architecture that uses AgentCore, Claude Sonnet 4.6, Titan Text Embeddings v2 and Stable Image Inpaint to generate PowerPoint files from enterprise documents. AWS reports a 70–90-second internal run and about $0.04 in model invocation cost, while identifying substantial supporting infrastructure and prerequisites.
  2. sample-generating-powerpoint-presentations-with-amazon-bedrock at commit 9c51d2cAWS Samples on GitHub · Reviewed: Public repository identity; MIT-0 license; pinned commit; README prerequisites; configuration example; CDK bootstrap, build and deployment steps; architecture note; default-branch metadata and latest commit returned by the GitHub API · Retrieved · Supports: The code artifact is public under MIT-0 and can be pinned to commit 9c51d2c881e9edbe81adbc84fedfa89cfa1b392b. Its README requires .NET, Python, AWS CLI, Node.js, CDK, an S3 bucket and an API Gateway timeout quota increase; its example model configuration differs from the October 14 blog post.
  3. Amazon Bedrock pricingAmazon Web Services · Reviewed: Current Stability AI Image Services on-demand table; Stable Image Inpaint per-generation price; managed Knowledge Base storage and retrieval pricing; exclusions and separate model-provider charges · Retrieved · Supports: AWS currently lists Stable Image Inpaint at $0.07 per generation in Oregon, N. Virginia and Ohio, and separately meters knowledge-base storage and retrieval. Those current list prices do not reconcile with the blog’s approximately $0.04 total model-invocation estimate for a presentation containing two generated images.
  4. Understand the AgentCore Runtime service contractAmazon Bedrock AgentCore Developer Guide · Reviewed: Runtime service-contract purpose; supported HTTP, MCP, A2A and AG-UI protocols; ports and mount paths; message formats; discovery; authentication options; documented use cases · Retrieved · Supports: AgentCore Runtime currently documents HTTP, MCP, A2A and AG-UI service contracts with distinct ports, paths and message formats. This establishes the managed runtime interface, not that the presentation sample is production-ready or portable without change.

Limitations

This is a desk review of AWS-controlled sources. AccessAllGPT did not receive AWS credentials or credits; deploy CDK stacks or containers; inspect every repository file; validate the repository’s build state; invoke AgentCore, Bedrock Knowledge Bases, Claude Sonnet 4.6, Titan Text Embeddings v2 or Stable Image Inpaint; upload the annual report; reproduce ingestion or 70–90-second generation; inspect the six-slide output; verify citations, quality, isolation, safety or deletion; load-test the system; or obtain an account bill. The timing and approximately $0.04 cost are AWS internal claims from one disclosed input. Current listed pricing conflicts with that cost statement, and service, model, region, price and repository contents may change.

Disclosures

AccessAllGPT received no AWS account, credits, code contribution, output, benchmark, bill, briefing, review, payment or compensation for this article. AWS and the sample authors did not sponsor, review or endorse it. AccessAllGPT did not test, score or rank the architecture or models. AccessAllGPT Research is operated by NeuralArc, is independent, and is not affiliated with AWS, Anthropic, Stability AI or organizations cited. Publication-wide relationships are listed on the disclosures page.

Further AccessAllGPT guidance

  1. Postman Details How Agent Mode Runs on Amazon Bedrock
  2. Third Moment Says It Runs 10,000 OpenAI Agents for Up to 48 Hours
  3. Design an Agent Benchmark That Predicts Production
  4. Control the Hand-off Between AI Agents and Human Approvers
  5. AI API Data Retention and Residency Procurement Gates
  6. The AI Tooling Procurement Scorecard
  7. AccessAllGPT Research methodology
  8. Evidence standards
  9. Publication disclosures

Continue the research

Get evidence-led updates for teams making production AI decisions.