Play the News · Drop #043 · Morning

Ship It or Block It

How it works: 6 cards, about 2 minutes, from this Drop's news. Your streak lives only on this device. It's not legal advice.

Card 1 of 6Agents
Your reinforcement learning run lets a frontier-class agent browse the live web freely to look up answers during training.
You're the one who signs off. What's the call?
Show the answer
Answer: Block it.
Sandbox isolation
OpenAI now trains frontier models on cached, indexed copies of the web instead of the live internet, TechCrunch and The Verge report. Two leading labs now say their strongest models can't be trusted online during training.
From the Drop: OpenAI cuts frontier model training off from the live internet
TechCrunch · Oct 23 The Verge · Oct 23 OpenAI · Oct 23 Hacker News · Oct 23
Card 2 of 6Jobs
A personal finance agent wants to move a customer's savings to a higher-yield bank on its own, without asking first.
You're the one who signs off. What's the call?
Show the answer
Answer: Needs a human.
Human authorization
Bank of America's CEO told CNBC a human should authorize money leaving an account, and JPMorgan says it blocks agents that rely on shared login credentials.
From the Drop: Bank CEOs push back on fears that AI agents will drain deposits, CNBC reports
CNBC · Oct 23
Card 3 of 6Agents
A troubleshooting agent reads a failed workflow's logs and status through a gateway with narrowly scoped tools, then explains the failure to the user.
You're the one who signs off. What's the call?
Show the answer
Answer: Ship it.
Least privilege
AWS says Bayer runs a Troubleshooting Agent like this in production among 22 specialist agents, cutting a diagnosis from over an hour to under 10 minutes, with tools scoped per agent.
From the Drop: Bayer's 22-agent decision engine answers about 80% of questions without an engineer
AWS Machine Learning Blog · Oct 23
Card 4 of 6Laws
Your product team wants to target ads using email content users connected to an AI assistant, without telling them.
You're the one who signs off. What's the call?
Show the answer
Answer: Block it.
Transparency and consent
Even Google's new policy ties ad use to data users choose to make available to AI features like Personal Intelligence, The Verge reports, and users can disconnect those apps.
From the Drop: Google's privacy policy now lets connected Gmail and Photos data shape ads
The Verge · Oct 23 Ars Technica · Oct 23
Card 5 of 6Agents
Your assistant learned a user wants no confirmations for work emails, and now also skips confirmation on personal payments.
You're the one who signs off. What's the call?
Show the answer
Answer: Needs a human.
Scope preferences
The MemPABench preprint's authors report that file-based memory improved accuracy but applied preferences beyond the contexts where they belonged.
From the Drop: Preprint: memory tools barely help agents learn how users want to be helped
arXiv · Oct 23
Card 6 of 6Agents
Your harness saves oversized tool results to files and shows the model a short preview plus the path, so it can read them back later.
You're the one who signs off. What's the call?
Show the answer
Answer: Ship it.
Recoverable context
LangChain says Deep Agents does this for tool results over roughly 20,000 tokens, arguing large inputs should be offloaded rather than truncated so nothing is lost.
From the Drop: LangChain details how Deep Agents offloads and summarizes context to fight context rot
LangChain · Oct 24

Games are written by AI from this Drop's sourced items. Scenarios are illustrative.