Experimental tool · rapplication

AI Review Walkthrough

Walk through a responsible AI review of your own AI agent or app, one stage at a time, right in your Brainstem chat. Each stage explains why it matters, asks a few plain questions, and shows a worked example from an invented sample agent. At the end you export a review packet and a one-page summary.

Built from Microsoft's public responsible AI guidance. Educational only. It is a generic review any team could run. Your answers are saved only in your Brainstem's local storage, and exports are written to a folder on your device. What you type in chat is sent to the AI model your Brainstem uses, as with any chat, so leave out anything you would not share there.

Experimental and not an officially supported Microsoft product or service. Provided as is under the repository's MIT license, with no support or warranty. Use at your own risk. Trademark use follows the repository's trademark notice.

Educational. Not legal advice, and not a substitute for your organization's own review.

Install into Brainstem

Run this once on the computer where your Brainstem runs. The agent is ready on your next chat message.

curl -fsSL https://microsoft.github.io/aibast-agents-library/docs/experimental/ai-review/ai_review_walkthrough_agent.py -o ~/.brainstem/src/rapp_brainstem/agents/ai_review_walkthrough_agent.py

Download the agent file instead, and drop it into your Brainstem's agents folder.

Use it in chat

The twelve stages

  1. Describe the systemWhat it does, who uses it, how it works and where it runs.
  2. Intended and out-of-scope usesWhat it is built and tested for, what it is not for, and its known limits.
  3. Stakeholders and affected peopleEveryone it touches, with likely benefits and harms, and who oversees it.
  4. Data and privacyWhat data goes in, whether it suits the use, and how it is stored and accessed.
  5. Harms mapping across the principle areasFairness, reliability and safety, privacy and security, inclusiveness, transparency, accountability.
  6. Threat model, in conceptWhat is valuable, where untrusted input comes in, and what protects it.
  7. Evaluation plan and defect ratesQuality, groundedness, harmful-content and misuse testing, against target defect rates.
  8. MitigationsWhat reduces each harm, in layers, with a fallback and proof it works.
  9. Human oversight and escalationWhere people review or override, and how to avoid over-reliance.
  10. Transparency noteA short note for users: what it is, its limits, that it is AI, where to get help.
  11. Release decision and sign-offGo, go with conditions, or not yet, signed off by named roles.
  12. Monitoring and incident responseWhat you watch after release, and how you respond, roll back and learn.

Sources

Public pages, readable without sign-in, checked 2026-10-29. The walkthrough paraphrases them in short plain words.