Field Notes

OpenAI sketches new agent stack, AWS shows practical RAG, no-code agents move in

Today is about agents, safe controls, and small steps you can use right away.

Key AI news

OpenAI sketches new agent stack: Decisions API, Agents API, and faster Computer Use

At DevDay, OpenAI showed new building blocks for agents: asynchronous tool calls, mid-turn steering (steer while the answer runs), WebSockets, UltraFast inference, and the Decisions API.1 Ari Weinstein said Computer Use is “180 degrees different” today, and screenshots plus accessibility data, DOM, Playwright, and generated code speed up tasks.1 Status: announced in the podcast conversation. Details and availability are vendor claims without independent verification.1 If you build agents, plan for structured decisions with low latency, plus longer prompt caches and pre-warming.1

White House accord: large labs commit to external audits and controls

A report says industry leaders signed a joint self-commitment with external audits, internal controls, and multilayer reviews. The US President called it “morally binding.”2 The list of signatories and the note that Microsoft and Amazon did not visibly sign come from this source and are not officially confirmed.2 If you deploy frontier models, prepare an external audit relationship and document your internal AI controls.2

ML early warnings for power grids during solar storms

Microsoft Research describes a system with 30–60 minutes of lead time and location-specific risks for 66,935 US substations.3 It detected about 80 percent of major events from 2020 to 2026. Reported are 76.5 percent hits at 10 nT/min, 81.2 percent at 20 nT/min, and 64.1 percent at 50 nT/min.3 If you work in grid ops or critical services in the US, feed such signals into readiness, alerting, and customer comms.3

Highlights for your workday

Build a claims-capable assistant with Amazon Bedrock Knowledge Bases: the Agentic-Retrieve-Stream API plans sub-queries, checks evidence, and returns answers with sources. A guardrail secures content binding to documents. Agentic retrieval means: the assistant plans and repeats fetches on its own.4 Google introduces Skills as reusable instructions and replaces Gems. You define your routine once and use it for recurring tasks.5 n8n shows a new agent with no code. Photos of sticky notes become tasks, receipts get logged, morning emails pre-sorted, controlled via Telegram.6

Tools and updates

  • LangChain Managed Deep Agents 0.8 now accept requests via arbitrary webhooks. An example logs expenses when you forward a receipt by SMS. Requirement: Managed Deep Agents access.7
  • Nemotron 3 Diarization in Transformers 5.18.0 separates speakers in real time and offline up to eight voices, ordered by arrival. Useful for meetings and call review. Speaker diarization means: “who spoke when.”8
  • NVIDIA NeMo Relay logs agent steps in detail so you can spot detours and cut latency and token use. Requires NeMo Relay.9

Try it in five minutes

1) Open your preferred AI chat and paste this sample text: “CLM-100482: auto claim, status open, filed on 07/09/2026, current amount 14,250 USD. The revised estimate from 07/22/2026 replaces the fax version from 07/15/2026. Provisional payment from 08/01/2026 was canceled on 08/12/2026. Planned check issuance after approval of the revised estimate.”4 2) Ask: “Was the estimate for CLM-100482 approved, and when will the check be issued? Answer only from the text and add a source bracket like [T1] after every sentence.”4 3) Check the answer. Ask the AI to fix it if it adds content not in the text, or if any sentence lacks a [T1] bracket.4

Expected result: a short answer with 1–2 sentences. Each backed by [T1]. No extra assumptions.4 Check criterion: every sentence relies only on the sample text and carries a source bracket.4

Sources

  1. Why Dwarkesh is Wrong about Computer Use + How OpenAI shipped its Jev competitor in 1 Week (latent.space)
  2. A ‘Morally Binding’ White House Accord on AI Safety (thezvi.substack.com)
  3. Forecasting space weather risks on power grids (microsoft.com)
  4. Query claims in natural language with Amazon Bedrock Knowledge Bases (aws.amazon.com)
  5. Let skills in Gemini tackle your most repetitive tasks (blog.google)
  6. @n8n_io: n8n now has a new kind of agent 🤖 Liam shows you how easy it is to set one up: a Telegram assistant that turns sticky-note photos into tasks, logs receipts, and sorts your email ev (x.com)
  7. @langchain: Managed Deep Agents 0.8 ships HTTP channels: agents can take requests from any webhook, not just Slack.

@ndrezn built an expense agent that lives in text messages.

Text a receipt (x.com) 8. Transformers 5.18.0 (github.com) 9. Tracing Agent Harness Behavior with NVIDIA NeMo Relay (developer.nvidia.com)