English

日本語

English

日本語

AI-Starter 3.1.8 ​

Last updated: October 22, 2026

AI-Starter 3.1.8 has been released. This update adds a new platform metadata type, "AnthropicMessages", which calls Claude through the Anthropic Messages API. It also improves how files are passed to Code Interpreter (Bedrock AgentCore) and makes its setup simpler. In addition, the conversation style (Temperature) setting, previously marked as deprecated, has been removed.

New Features ​

Calling Claude Through the Anthropic Messages API ​

A new platform metadata type, AnthropicMessages, has been added. It calls Claude on Amazon Bedrock through the Anthropic Messages API instead of the Converse API. Calls are made with InvokeModel on the bedrock-runtime endpoint. You can configure the max tokens and the thinking effort level. Adaptive Thinking is always enabled, and a summary of the thinking is displayed.

A new default assistant, "Claude Haiku 5.5 (Messages API)" (jp.anthropic.claude-haiku-5-5), has been added. It is disabled by default, so enable it in the admin console if you want to use it.

AnthropicMessages has the following limitations.

  • Tool calling: Not supported (including the BedrockClaude tools and MCP servers). To use tools, use BedrockClaude
  • Attachments: Images, PDFs, and text files only. Spreadsheets (XLSX), DOCX, and other formats are not supported
  • Supported models: Models that support Adaptive Thinking (Claude Opus 4.6 / Sonnet 4.6 or later)

For a full list of supported models, see Models.

Improvements ​

Improved File Handling in Code Interpreter ​

The way Code Interpreter (Bedrock AgentCore) handles attached files has been reworked.

  • Attachments are copied directly into the sandbox: The AI-Starter server now writes attachments directly into the inputs/ folder in the sandbox. The AI can read files without going through S3, which makes processing more reliable
  • Generated files carry over: Files generated earlier in the same chat are restored into the outputs/ folder of the new session
  • Expired session detection: When the stored session has expired, a new session is now started automatically
  • Large file handling: File transfers have been adjusted to stay within the AgentCore payload quota. Files that are too large are not transferred

Code Interpreter sessions time out after 15 minutes. Previously, files generated earlier could no longer be accessed once a new session started. Starting with this release, you can continue your work with previously generated files even after a break.

Simplified Code Interpreter Setup ​

With the improvements above, the AgentCore Code Interpreter resource no longer needs an IAM role (execution role) for S3 access. If you are setting up Code Interpreter for the first time, you no longer need to create an IAM role in advance. For the steps, see AgentCore Code Interpreter.

Environments that already use an execution role continue to work without any configuration changes.

Easier Editing of Long Prompt Templates ​

Fixed an issue where the save button in the prompt template editor could be pushed off screen when the content was long. The input field now has a maximum height, and long content can be scrolled and edited within the field.

Security Enhancements ​

Dependency packages have been updated to strengthen security.

Changes and Deprecations ​

Removed the Conversation Style (Temperature) Setting ​

The conversation style (Temperature) setting, previously marked as deprecated, has been removed in this release, because temperature cannot be changed on many of the latest models. Responses are now generated with each model's default value.

  • Chat screen: The "Conversation style" selection before starting a chat has been removed
  • Admin console: The "Conversation style (temperature)" assistant setting and the "Search query generation temperature" setting of the Bedrock knowledge base grounding have been removed

Temperature values stored in existing assistants and chat history are no longer used. To adjust the tendency of responses, describe the desired response style in the system prompt. For details, see Chat.

Closing ​

Thank you for reading. AI-Starter continues to evolve based on your feedback. If you have any questions or suggestions, please don't hesitate to let us know. Stay tuned for future updates!