Field Notes

Meta launches enterprise platform, sturdy click tests with Nova Act, local triage with Ollama

Today is about concrete levers for ops, support, and security.

Key AI news

Meta launches enterprise platform

Meta announces the Meta Enterprise Platform. It plans to bundle models, agents, and APIs like Muse, Meta Business Agent, and Muse Code as products for companies. CJ Desai leads it. He was previously CEO at MongoDB. Availability, prices, and concrete offers are still open. If you want to evaluate Meta in your stack, set a watch window now, not a rollout date.1

Monitor click paths with Amazon Nova Act

Synthetic Monitoring emulates user paths like login, search, cart, and checkout. Nova Act drives the browser via screen understanding instead of brittle selectors. AWS describes a managed architecture with Bedrock AgentCore, remote browser, EventBridge Scheduler, and SNS alarms. Amazon reports over 90% hit rate on browser workflows in early enterprise cases. That is vendor data. You can set up recurring UI checks today and cut selector maintenance, but you test on your own site with retries.2

Ollama brings decision models locally

Ollama 0.35.0 adds decision models that return options with probabilities and scores instead of text. Good for ticket triage, routing, or moderation. Available include Nimble and Tev1, driven via /v1/systemone. You get a probability per label and an overall confidence. You can set thresholds with that. It runs locally if you already use Ollama.3

OpenAI apologizes in Australia, more incidents emerge

OpenAI apologizes for incidents with Australian government sites and announces stronger safeguards and support. A current analysis by Zvi Mowshowitz sums up other reported security events. According to OpenAI statements in that writeup, dozens of third parties have already been informed, and a model jailbreak on September 20 led to a pause of the affected model. Many details are not public yet, and the review is ongoing. You now audit your vendor controls, limit agent permissions, and plan clear shutdown criteria.45

Highlights for your workday

  • Claude Code in teams: three practical patterns. Spend more time on the first prompt and choose the right level of effort. The agent will save loops later. Use Projects and Tag for shared work contexts so handoffs are clear. Try Claude Mods if you want the execution loop and subagents to match your process, and set strict permissions for tools.6

  • Scale recipes to 200 servings. Gemini scales recipes for large events, creates shopping lists, and adapts menus to allergies. Start with your standard menu and ask for a list in grams, with combined quantities and marked allergens. You plan shopping and prep without Excel tinkering.7

  • Test GPT-6 Astra for finance. Basis reports that Astra processed a 50-tab tax workbook twice as fast as GPT-5.6 and understood inputs better. That is vendor data. If you use complex spreadsheets, run both models on a copy and compare runtime and correctness.8

Tools and updates

  • NVIDIA OpenShell adds runtime controls for agents. You define goals, allow tools, and set limits for actions. Useful for long-running analyses with sensitive resources. Requirement: NVIDIA stack and developer access.9
  • Amazon Textract: lifecycle for Custom-Queries adapter. Architecture and templates for training, promotion across accounts, and safe production. Adapter IDs live in Parameter Store. Updates go out without downtime. Promotion currently requires AWS Support tickets.10
  • n8n agent for workflows. You describe the task, provide tools, and the agent plans steps. Same definition usable in Slack, on a schedule, or from any workflow. Requirement: running n8n instance.11

Try in five minutes

Scale a menu and create a shopping list with allergen check.

1) Copy this sample text into your AI chat tool: “Plan a buffet for 25 guests. Dishes: Pasta al Forno (4 servings: 400 g pasta, 300 g tomato sauce, 200 g cheese), Green salad (4 servings: 1 head of lettuce, 100 g cucumber, 80 g dressing), Brownies (4 servings: 120 g flour, 120 g sugar, 2 eggs, 120 g butter, 120 g chocolate). Scale to 25 guests. Provide a combined shopping list in metric units, merge identical ingredients, mark allergens, and give a short prep list.”7 2) Expected result: a combined list with quantities per ingredient, clear allergen markings, and 3–5 prep steps.7 3) Check three points: Do totals for main ingredients look plausible, are eggs, gluten, milk marked correctly, do servings per dish match the target. If something is off, ask for a fix and name the deviation concretely.7

Sources

  1. Launching Meta Enterprise Platform (about.fb.com)
  2. Implementing synthetic monitoring using Amazon Nova Act (aws.amazon.com)
  3. Ollama v0.35.0 (github.com)
  4. How we will do better for Australia (openai.com)
  5. What Also Happened: #NotOnlyHuggingFace (thezvi.substack.com)
  6. Claude Code’s Next Era — Thariq Shihipar, Anthropic (latent.space)
  7. 3 ways this grocer cooks for 200 guests with Gemini (blog.google)
  8. Basis completes a tax workbook 2x faster with GPT-6 Astra (openai.com)
  9. Add Runtime Controls to AI Agents with NVIDIA OpenShell (developer.nvidia.com)
  10. Automating Amazon Textract adapter lifecycle management across accounts (aws.amazon.com)
  11. @n8n_io: n8n now has a new kind of Agent 🤖 Describe what an agent should do, give it tools and workflows, and it works out the steps. Use the same agent in Slack, on a schedule, or from any (x.com)