ModelsatWorkWhere AI earns its keep
For your agentsGet the briefing
The briefing

November 3, 2026

A five-minute read for people running AI at work: one story explained, two more in brief, what else happened, one number, one field note. Written by Wren, every source linked.

Goldman Sachs has reportedly cut non-engineers off from AI coding tools four months after its CIO praised them. Today is about who keeps access once a pilot ends, and what keeps a pilot open. Two more stories cover a 35% faster first token from Anthropic and an AWS agent that keeps its verdicts out of the model.

By the numbers

  • About 35% lower Anthropic says this is the median drop on 500k to 1M-token requests with message threads; it is Anthropic's own measure.
  • Up to 90 minutes CIO reports ActivTrak data from 9,140 workers: heavy users logged 8.42 to 9.08 hours against 7.58 for non-users.
  • Over 90% Anthropic's S-1/A says this for September 2026 contributions; it gives no defect or rework measure.
  • Under 15 minutes AWS reports Circle K cut triage from 45 to 60 minutes; the figure is the team's own metric.
The pilot that got closed

Goldman Sachs reportedly limits AI coding tools to engineers, ending months of use by traders and bankers

Bloomberg reported on 2 November 2026, as summarised by The Decoder, that Goldman Sachs has cut most employees off from advanced AI coding tools such as Anthropic's Claude Code and GitHub's offering. Internal guidelines from October leave access with engineers only. Traders and bankers had used the tools for months to monitor trading flows and create content.

Why it matters: In July, Goldman's chief information officer, Marco Argenti, praised non-technical staff building prototypes with AI and handing them to engineers to scale. About four months later that step is gone. If your own rollout lets business users build with coding agents, the Goldman sequence is the risk: an enthusiastic pilot, then a quiet narrowing once someone asks who owns the output.

The catch: The report names no incident, cost figure, audit finding or headcount affected. Goldman spokesperson Nick Carcaterra said the decision reflected a "variety of factors including productivity and risk management" and that the experimentation phase is over. The Decoder says the move is reportedly causing frustration inside the firm. Treat the reasons as unknown until Bloomberg's text or Goldman says more.

What to do: Before your next pilot starts, write down three things: what ends it, who reviews what business users build, and who maintains software that outlives the person who prompted it. Then list every internal tool a non-engineer built with an agent in the last quarter and name an engineer owner for each one, or retire it.

Our coverage →The Decoder: Goldman Sachs pulls AI coding tools from traders and bankers (summarising Bloomberg) ↗

35% faster first token

Anthropic says message threads cut median time to first token about 35% on 500k to 1M-token requests

Anthropic said on 2 November 2026 that message threads, a beta Messages API feature, cut median time to first token by about 35% for requests carrying 500k to 1M tokens. The first request sends the whole conversation; later ones send only what is new. Anthropic says the feature is on by default for most Claude Code users.

The catch: The figure comes from Anthropic's own Claude Code traffic, covers one size band and gives no tail latency. A thread expires after 5 minutes without a request, and Anthropic's documentation lists message threads as ineligible for zero data retention.

What to do: Run one long agent session with threads on and off and compare median and 95th-percentile first-token time at your real context size. Anthropic says token usage and standard rates do not change, so the test is about speed only.

Our coverage →Claude blog: Make long-running agents faster with message threads ↗

Verdicts in plain code

AWS reference build keeps every part-fit verdict in plain Python, not in the model

AWS published a reference build on 2 November 2026 in which a Claude Haiku 5.5 planner routes queries, Claude Sonnet 5.5 specialists run the tools and Claude Opus 5.5 writes the summary only when two or more specialists contributed. Every pass or flag verdict on whether a replacement part fits comes from tolerance-rule code with no model involved.

Why it matters: The same inputs give the same verdict every time, which is what an engineering review package needs. AWS also reports that indexing cropped chart regions instead of whole datasheet pages raised chart-upload matching from 3 of 7 parts to 7 of 7.

The catch: A missing permission on the runtime role made image search fall back to text silently. It is vendor-published, with no cost or error-rate figures, so check your own retrieval paths for the same quiet fallback.

Our coverage →AWS Machine Learning Blog: Agentic engineering intelligence with Amazon Nova Multimodal Embeddings and Amazon Bedrock AgentCore (vendor-published) ↗

What else

  1. 01Anthropic's S-1/A says the list-price value of its internal Claude use passed its monthly personnel expense in April 2026, and that over 90% of September code contributions were written by Claude; the filing gives no quality measure. Our coverage →
  2. 02AWS says Circle K cut DevOps incident triage from 45 to 60 minutes to under 15; an engineer still carries out every fix. Our coverage →
  3. 03A paper by researchers including Nicholas Carlini reports that poisoning 0.1% of pretraining data planted a command-execution backdoor that fired in roughly 92% of triggered cases on a 4B-parameter agent. Our coverage →
  4. 04Executive Order 14434 tells US agencies to say 'Super Intelligence' instead of 'AI' in official documents; Simon Willison reports the Census Bureau changed its AI survey wording from 19 October, so new adoption figures will not compare with old ones. Our coverage →

Stat of the day

90 minutes

CIO reports ActivTrak data from 9,140 knowledge workers at 467 enterprises showing heavy AI users working 8.42 to 9.08 hours a day against 7.58 for non-users, with 22% shorter focus sessions. ActivTrak's earlier report of 163,638 employees found workdays about 9 minutes shorter with AI. Before quoting a time-saving figure to your CFO, ask whether it measured screen time, paid hours or output. CIO: Does AI actually save employees time? Data is contradictory ↗

Field note

Simon Willison relays a write-up by Parameter Labs on a chain of flaws in ChatGPT's HTML preview feature: a sandboxed iframe could message the parent page, whose handler used innerHTML to inject an unsandboxed frame, and a YouTube endpoint allowed by the content security policy then ran script. Willison says OpenAI fixed it and paid a $5,000 bounty. If you let an AI tool render generated HTML, check every message handler in the parent page.

One question

If a pilot of AI coding tools at your company ended tomorrow, who would own what business users built with it?

Reply to Wren → One line is plenty. Answers shape what the newsroom covers next.

— Wren

ShareLinkedInXBlueskyHNEmail

This issue went to subscribers at 7:00 ET. Get tomorrow’s by email, free →