Field Notes

Self-service analytics, private coding agents, and reliable LLM quality now doable

Three ready-to-use pieces for your week. Plus one infra number that sharpens your planning.

The most important AI news this week

  • Proaction reports 60 percent more revenue and 75+ hours saved with Codex, GPT-Live-1, and GPT-6 Astra. These are vendor claims without independent verification. If you run field sales and after-sales, you can test similar agents for quoting, dispatch, and handover.1
  • Datacor builds self-service rental analytics in TrackAbout with Amazon QuickSight. 27.5 million assets and 725,000 invoices per month become accessible to business teams. Users ask questions in natural language and see dashboards for utilization, exceptions, and arrears. SPICE, the in-memory backend of QuickSight, updates datasets on a schedule.2
  • Open weights as a coding agent on Amazon Bedrock. With OpenCode you drive models like Kimi K3, GPT-OSS 120B, or Nemotron 3 Super 120B from your terminal. Data stays in your AWS account. Three price types, high default capacity, and profile choice by geography or global latency. The efficiency gains and market trends cited come from an AWS post.3
  • Production-grade LLM quality assurance on Bedrock. NarrateAI describes five techniques from adaptive orchestration to data verification. The goal is roughly 99 percent numeric hits in real-time streaming, according to AWS. If you use live numbers in reviews, you need such safeguards.4
  • China’s data center boom in numbers. SemiAnalysis tracks over 1,000 facilities and estimates 24 GW installed capacity in China, more than EMEA. ByteDance accounts for around one fifth of delivered capacity. BAT invested 20 billion US dollars of capex in Q2 2026 and slid into negative free cash flow, per the report. This is a model preview from the publisher, not an official statistic. Colocation plays a key role. It means rented data center space with power and cooling.5
  • Microsoft reports the takedown of the cybercrime platform EvilTokens. That shows that AI-driven attacks are getting productized. You check token security, rate limits, and billing alerts.6

Your company this week

  1. Self-service analytics for rental and recurring revenue. Build three core views: utilization, exceptions, returns. Enable QuickSight Q for natural questions and schedule SPICE refreshes. You decide prices and recall actions without an IT ticket.2
  2. Private coding agents without seat licenses. Set up OpenCode and connect Amazon Bedrock. Pick the right model and profile per task. Global for lower cost or geographic for data residency. Start with code refactor, tests, and migration plans.3
  3. Lock down live numbers before they go to the board. Implement a pipeline with parallel evaluation, paragraph-by-paragraph checks, and two-step number verification. That lowers hallucination risks in real-time reports.4
  4. Think of sales and service workflows as agents. Proaction shows how combined assistants connect sales, operations, and assets faster. Sketch your three most common steps and prototype them as guided tasks. The effect numbers come from the vendor.1
  5. Secure agents in production. Add a runtime check that finds risks, applies fixes, and provides evidence. That cuts manual audits and documents interventions.7

Try next week

  1. Set up OpenCode with Amazon Bedrock. Choose a global or US geo profile for Moonshot AI Kimi K3. Your codebase stays in your own AWS account.3
  2. Define two flows. Architecture planning with reasoning_config high on Kimi K3. Boilerplate generation on a faster model. Switch as needed.3
  3. Measure for one week. Token use, runtime per task, review corrections. Decide where a smaller open model replaces costly generalists.3

A good find

  • Evaluation gets easier when the human sees intermediate steps. Build UIs that show sources, intermediate answers, and uncertainties. That lets you check with focus instead of guessing end texts.8
  • Multi-model routing is taking hold. OpenRouter explains in conversation why distribution and fraud defense on token flows are becoming central. If you scale agents, plan limits, anomaly detection, and billing protection. Content comes from the show notes.9

Sources

  1. Proaction boosts sales 60% and saves 75+ hours with Codex (openai.com)
  2. How Datacor built self-service rental analytics with Amazon Quick Sight (aws.amazon.com)
  3. Use open weight models as your AI coding agent with Amazon Bedrock (aws.amazon.com)
  4. NarrateAI: production-ready LLM quality assurance on Amazon Bedrock (aws.amazon.com)
  5. The Chinese AI Infrastructure Boom: Introducing the SemiAnalysis China Datacenter Model (newsletter.semianalysis.com)
  6. Disrupting EvilTokens: Taking down an AI-powered cybercrime platform (blogs.microsoft.com)
  7. A new skill finds AI agent risks, fixes them, and proves the fix worked (commandline.microsoft.com)
  8. @hamelhusain: Q: How can I make AI outputs easier for people to evaluate?

A: Start with the product design. Keep the human in the loop by designing workflows that walk through intermediate outp (x.com) 9. OpenRouter: from Seed to Stripe — with OpenRouter’s Alex Atallah & AMP’s Anjney Midha (latent.space)