Field Notes

Support triage, call center agents, and Muse on glasses: three quick levers for your day

Today is about back-office automation, customer service costs, and AI at your eye on the go.

Important AI news

Muse is coming to AI glasses, plus new connectors and email

Meta is bringing the personal agent Muse to AI glasses in the coming months. You say the name and Muse can act based on what you are looking at. That saves explanation and speeds up tasks on the go. New are many commerce and work connectors, including Walmart, Best Buy, Sephora, Wayfair, Notion, GitHub, and Box, plus Shop Pay and PayPal. Muse also gets its own email address for mail-based workflows. This is announced, not live. Vendor claims.1

Support triage with Amazon Nova: 96 percent routing hits at Aderant

Aderant automates ticket pre-sorting with Amazon Nova Lite via Bedrock. The analyzer checked 109 tickets in 2.5 weeks and reached about 96 percent routing accuracy. Estimated relief: 8 to 14 engineer hours per week with under 30 dollars monthly operating cost, of which under 1 dollar for inference. Low confidence goes to human review. The accuracy was measured against the team that solved the case in the end.2

Ringg: AI agents resolve up to 65 percent of calls

Ringg uses GPT-5.6 for multilingual agents across phone, chat, WhatsApp, and web. The vendor cites up to 65 percent resolved calls and 90 percent lower costs versus GPT-4.1. These numbers come from the vendor and are not independently verified.3

Open 3D structures for 2,800 viruses in the AlphaFold database

Google DeepMind, EMBL-EBI, and NVIDIA release predicted protein complexes for over 2,800 viruses. Around 30 percent of the interactions are newly described. The underlying workflow, the BioNeMo Structure Prediction Pipeline, is provided openly. This helps research teams narrow targets faster. For most companies this is more like foundational work.4

Runway’s WorldPrompt: control worlds in real time, as a research preview

Runway shows a format proposal called WorldPrompt with GWM Worlds 2. You set a start image and time-stamped events and control an interactive, real-time simulation. This is a research preview with limits on reliability and duration. The basis is podcast show notes, not the paper itself. Uses: trainings, demo experiences, prototypes.5

Highlights for your workday

  • You can automate ticket triage step by step. Start with hourly runs on weekdays, clear confidence thresholds, and automatic routing only at high certainty. Build first answers from Confluence, SharePoint, and similar sources right into the analysis.2

  • You can absorb call spikes. Deploy a voice agent first for opening hours, order status, and simple callbacks. Measure resolution rate, escalation rate, and cost per contact. Vendor figures of 65 percent completion are not independently proven.3

  • You can fine-tune your own classifier for 17 dollars. Together shows step by step how to train a Jev-like filter on a 4B model. Use cases: lead qualification, spam filtering, routing of forms. Prices and effort are vendor claims.6

Tools and updates

  • LangSmith expands fine-tuning and monitoring. You can train specialized models from real usage logs and use Engine v2 to spot agent issues proactively. Public beta. Requirement: existing traces and existing agent workflows.78

  • NVIDIA Confidential Computing for private inference. LLM workloads run in memory-encrypted environments on confidential VMs and GPUs. Interesting for teams with sensitive data and compliance needs.9

  • Together: “Train your own Jev” guide. Serverless fine-tuning of a 4B classifier on Together with sample code. Useful when you need low-cost domain-specific filters.6

Try in five minutes

Test pre-sorting of support tickets in chat.

1) Copy the five sample tickets into your AI chat and ask for routing to teams: CloudOps, App Support, Billing. Ask for confidence per ticket and a first step to solve as bullet points.

Sample tickets: - “Invoice 84721 shows duplicate line items for August. Customer ACME GmbH.” - “Deploy of release 3.2 fails in prod-eu-1, code 503 at /health.” - “User cannot log in since yesterday, password reset does not help.” - “S3 backup job red since Monday, 0 bytes copied, region eu-central-1.” - “Customer asks about upgrade discount when switching from Standard to Pro.”

2) Ask the model: “Route only automatically if confidence ≥ 0.8. Else send to manual review.” Have it output the decisions as a table.

Expected result: sensible assignment of tickets to the three teams plus a concrete first step for each case. Check: at least four of five tickets assigned correctly, confidences justified plausibly.2

A good find

Foundries vs. Navigators in day-to-day biotech. The piece shows how labs ship more experiments with quick internal tools, even without big facilities. First: analysis code adapts to new protocols in hours instead of weeks. Second: own data portals mirror the team’s view, not that of a standard tool. Third: decisions get broader because more candidates get pre-screened fast. Apply the principle. Build a lightweight internal dashboard for your team that answers your exact questions in one click.10

Sources

  1. The Biggest News From Connect 2026 (about.fb.com)
  2. Aderant builds intelligent ticket triage with Amazon Nova (aws.amazon.com)
  3. Ringg’s AI agents resolve up to 65% of customer calls with OpenAI (openai.com)
  4. How Open Science Can Help Researchers Prepare for the Next Pandemic (blogs.nvidia.com)
  5. Runway’s WorldPrompt and the Engineering of Real-Time Worlds (latent.space)
  6. How to train your own Jev for $17 (together.ai)
  7. @langchain: Introducing LangSmith Fine-Tuning and the smithtune CLI.

LangSmith now handles the entire fine-tuning process. Use your traces to train specialized models that cut cost and latenc (x.com) 8. @langchain: Just announced at Interrupt: LangSmith Engine v2.

With Engine v2, you can proactively spot agent issues and resolve them before your users see them.

Your agents get better. You g (x.com) 9. Enabling Private High-Performance Production AI Inference with NVIDIA Confidential Computing (developer.nvidia.com) 10. Foundries vs Navigators: Lowering the Cost of Science (latent.space)