Field Notes

Anthropic checks from the inside, Meta protects glasses data, Astra cases show productivity gains

Today is about audited models, confidential processing in the cloud, and proven efficiency jumps.

Big AI news

Anthropic launches embedded audits with Accenture

Anthropic and Accenture agree on a partnership for embedded auditors. These are independent teams inside the company with internal access. Both sides plan at least 1 billion dollars each over five years, according to the announcement. The collaboration is not exclusive. METR is set to pilot elements as well. Anthropic funds the Accenture work directly, with more evaluators to follow. For you this means more audit paths and outside eyes on frontier models. Even more so if you use Anthropic.1

Meta brings Private Processing to AI glasses

Meta describes Private Processing for AI glasses. It is confidential cloud processing on hardware-protected environments. The approach extends the 2025 tech to glasses and aims to process personal context without Meta seeing data in operation. Execution runs in isolated virtual machines with attestation, according to Meta. If you work with glasses, you can plan context-rich features and address privacy needs at the same time.2

Astra in practice: law, research, and video

New case studies show GPT-6 Astra in companies. Harvey reports more structured legal drafts. Parallel says time and cost are halved for labor-market research. invideo reports three times faster color correction and 50 effects in one day. Higgsfield ships video features within a day. These are vendor claims without independent review. If you have similar tasks, try a limited pilot with clear metrics.3456

New open MoE model: Hunyuan-A13B

Hunyuan-A13B is an open mixture-of-experts model with 80 billion total parameters. 13 billion are active at runtime. The team cites 20T training tokens and focuses on factuality and reasoning at lower operating costs. Interesting if you want to evaluate a capable, more efficient open model. Check license and terms before use.7

Highlights for your workday

Make unstructured media usable in Salesforce. AWS shows how to connect Agentforce with Amazon Bedrock Data Automation and the Model Context Protocol. Upload, auto extraction, and querying happen from within the Salesforce chat. Next step: Check if your edition allows external MCP servers.8

Build an assistant with device storage. DeepLearning.AI offers a short course with Qdrant on local storing, search, and forgetting of text and image memories. You learn to set up an offline assistant for Mac, Windows, or Linux without sending data to the cloud. Useful if privacy matters to you.9

Evaluate, do not guess. Hamel Husain advises an eval audit to find quick error sources. NVIDIA explains how to grade agents from tool call to task completion, not just pretty answers. Start with a real end-to-end task and measure completion rate and retries.101112

Tools and updates

  • Dify Agent: Build agents by conversation, test in preview, and roll them out to the team. Good if you want one surface from ideas to operations. Requirement: Willingness to introduce new agent workflows.13
  • LangSmith Custom Apps: Generate interfaces from agent data with a prompt. Useful to sketch internal dashboards without a frontend team. Requires LangSmith.14
  • Gemini Connected Apps: New links to Adobe, Airtable, Linear, Peloton, and more. You can trigger tasks across connected tools directly from Gemini. Requirement: App links enabled in your account.15
  • Replit for Meta devices: Describe a VR app, Replit creates the starting point for Meta devices. For teams that need quick prototypes in VR. Access to Replit and Meta target devices required.16

Try in five minutes

Your own practice idea based on spec-driven development: write a mini spec and let the coding agent implement it.17

1) Copy this spec into your coding model: "Build a CLI tool in Python that reads a CSV with columns name, qty, price and outputs a JSON list with fields name, qty:int, total:float. total = qty*price, two decimal places. Skip bad rows and count them. Output: JSON and at the end 'skipped: X'." Example CSV:

name,qty,price Pen,3,1.20 Notebook,2,5.00 ErrorRow,abc,9.99

2) Have the agent deliver the code and run it locally or in your tool's browser interpreter. Use the example CSV as input.

3) Check criterion: The JSON output contains two objects, totals 3.60 and 10.00, and the last line reads "skipped: 1".

Expected result: A short, testable script that follows your spec and handles bad rows cleanly.17

A good find

Practical AI on agents in the enterprise: bring platform engineering. Define identity and rights for agents separate from the user. Enforce sandboxes and put an MCP gateway between chat and tools. Keep shared agent memory small and auditable. This helps you firm up security and the path to production before you scale.18

Sources

  1. The Quest for Embedded Evaluators (thezvi.substack.com)
  2. Bringing Private Processing to Meta AI Glasses (engineering.fb.com)
  3. Harvey turns legal context into stronger drafts with GPT-6 Astra (openai.com)
  4. Parallel cut research time and cost in half with GPT‑6 Astra (openai.com)
  5. How invideo improves color grading 3x with GPT‑6 Astra (openai.com)
  6. Higgsfield AI ships new video features in a day with GPT-6 Astra (openai.com)
  7. Paper: Hunyuan-A13B Technical Report (huggingface.co)
  8. Extending public sector intelligence with Agentforce and AWS (aws.amazon.com)
  9. Building AI Assistants with On-Device Memory (youtube.com)
  10. @hamelhusain: I recommend grabbing these skills that @sh_reya and I created and at the very least doing an /eval-audit of your existing pipeline.

We've found that people often find low hanging (x.com) 11. How to Evaluate AI Agents From Tool Calls to Task Completion (developer.nvidia.com) 12. @hamelhusain: This post was a labor of love! We’ve distilled thousands of hours of work on AI Evals into a 30 min read + skills you can use to quickly uncover errors in your product.

https://t. (x.com) 13. @dify_ai: Missed the new Dify Agent launch? Here’s the experience in two minutes.

Start with an idea, build through conversation, and test in Preview. Build once, then use it across the pla (x.com) 14. @langchain: Introducing LangSmith Custom Apps

Create any interface from your agent data with a prompt. If you can think it, LangSmith can build it.

Now GA. https://t.co/MayBpUbPdu https://t. (x.com) 15. A new wave of Connected Apps is rolling out to Gemini. (blog.google) 16. @replit: Your next app could be in VR.

Describe an app, and Replit builds it for @Meta devices.

Announced today at #MetaConnect. Start building: https://t.co/NIaZFInsRi https://t.co/Hmgc1 (x.com) 17. Full Course: Spec-Driven Development with Coding Agents (youtube.com) 18. From AGENTS.md to Enterprise Deployment (share.transistor.fm)