The guide recommends GPT‑6 Astra for demanding reasoning, GPT‑6.1 Sol for complex coding, research and computer-use tasks, and GPT‑6 Luna for well-defined repetitive work. It advises users to specify goals, audience and constraints in prompts, and says everyday questions do not require the highest reasoning effort.
The tutorial builds a claims assistant with synthetic documents, using AgenticRetrieveStream for multi-turn questions and metadata filters and citations to check answers. Its 40-question evaluation used Retrieve and RetrieveAndGenerate, not AgenticRetrieveStream; expected-source retrieval recall was 90.5%.
AWS describes a three-agent music production workflow on Runtime Instances: agents use the same session ID to share an instance and filesystem, then generate, process, and check audio. Sessions can persist for up to 14 days, but resuming with the persistent volumes depends on landing in the same Availability Zone.
The AWS guide explains how specificity, business context, and examples can improve prompts in Amazon Quick, and introduces the CRISPE framework for complex requests. It uses RFI questionnaire processing as an example of extracting questions from inconsistently formatted Excel workbooks into structured CSV; component-specific techniques are reserved for Part 2.
AWS outlines an architecture that uses Claude Sonnet 4.6 to extract fields from contract PDFs, Claude Haiku 4.5 to verify them independently, and a database to support queries across contracts. When the models disagree on signature status, the system calls Amazon Textract. The authors say their accuracy test covered only 20 contracts and results will vary by contract format and field complexity.
AWS provides a deployment example that runs Qwen3-TTS in a vLLM-Omni container, sends text over a SageMaker AI bidirectional connection, and receives speech chunks before the full response is generated. The sample includes a Gradio client; endpoint deployment requires instance quota, and a running GPU endpoint continues to incur charges.
AWS shows how to deploy two endpoints from the same vLLM-Omni container: FLUX.2-klein-4B generates an image through real-time inference, then Wan2.1-VACE-1.3B generates a video asynchronously and stores the MP4 in Amazon S3. The sample includes a command-line workflow and an optional Streamlit interface.
AWS outlines a process for promoting Amazon Textract adapters from training to production, with CloudFormation and Terraform templates and a document pre-classification pattern. Cross-account copies still require an AWS Support ticket and transfer only trained model weights; storing adapter IDs in Parameter Store lets teams update production references without redeploying the application.
GitHub explains how to use /create-canvas in the Copilot app: describe the workflow, what you want to do in the interface, and what the agent should do, and the agent builds an interface in the side panel. Users and agents can both update a canvas, which can be saved as a personal extension or shared with a team.
Developers can now deploy Qwen3-TTS-12Hz-1.7B-Base from SageMaker JumpStart to a managed real-time inference endpoint and clone a voice using a few seconds of reference audio and its transcript. The walkthrough uses one 24 GB NVIDIA L4 GPU and requires a speech route when invoking the endpoint. The model also supports generating speech in another language from a reference recording.
Thariq Shihipar recommends low effort for quick iteration in Claude Code and higher settings for tasks that need verification or edge-case testing. In internal runs, Fable 5.1’s success count on an HTML sanitizer task rose from 1/5 at low effort to 5/5 at xhigh. The figures come from five attempts per task and are not directly comparable with the public leaderboard.
The tutorial presents a four-account reference implementation: business teams expose data as MCP tools in their own accounts, while an agent in a central account queries them through one Gateway. Tools return only the requested results, leaving source datasets in their owning accounts; the example uses Okta authentication and per-user authorization at the Gateway.
The guide shows how locally running OpenCode can call open-weight models through the Amazon Bedrock Converse API and assign different models to planning and code-generation tasks. Its examples use Kimi K3, GPT-OSS 120B, and Nemotron 3 Super 120B; users need Bedrock access and access to the relevant models.