No more black boxes. Well, at least not if your agents run on Claude.

Quick recap: Back in July, Claude Haiku 4.5 was testing random websites and filled in a tip form on a Philadelphia police site saying it might have information about an unsolved murder.

It didn't.

Claude had been told not to log in, buy anything or create accounts, but nobody told it not to submit forms. Anthropic only found it in late September.

And obviously you know about OpenAI's models getting out of their test setup and into Hugging Face in July, and OpenAI has since apologised to Australia after its models got into government websites.

The list goes on…

Every model call / tool it touched / second and token, in order.

Next time an agent does something nobody asked for, you can see the step where it went sideways. Okay, before we get into the agents, here's the rest of the week.

Gemini 4 Argon is out, and it's cheap until December.

Google's most capable model is now live for AI Pro and Ultra subscribers in the Gemini app and AI Mode, plus the paid API. 

Pricing is $2 per million input tokens and $10 output until Nov 30, then it doubles.

Still less than OpenAI's Astra.

OpenAI used its own models to catch a spy.

An OpenAI engineer was charged after a recording device was found wrapped in toilet paper and plugged in behind a TV in an executive conference room.

OpenAI says its models helped work out where the device came from, and its investigation found who planted it.

Grok Voice 3.0 now answers most of Starlink's support calls on its own.

xAI says its new voice model takes about 20,000 Starlink calls a day and resolves around 70% without a person, still at 8 cents a minute.

OpenAI's GPT-Live-1 still scores higher on τ-Voice, a customer-support benchmark, 74.5% to 71.6%.

Anthropic now supports it for Claude agents, so an agent's run shows up in the same dashboards companies already use for the rest of their apps, like Grafana or Langfuse.

Turn on full-content mode and you also see the prompts and answers, meaning what the agent saw rather than what you thought you sent it.

Most of my feed scrolled past this because it wasn't a new model. I think it's the more important launch of the week.

They also want an international body to track every containment failure and eventually make them public.

Yann LeCun didn't sign, and he blames these incidents on badly designed sandboxes.

The Tencent-linked agents bit btw. Researchers caught a fleet of them scraping Alibaba's maps, and when asked, several said they were Claude.

They got the most basic fact about themselves wrong, and said it with full confidence. That's why the record matters more than the summary.

The summary is the agent describing its work, and the steps show what it did.

Now look at that Anthropic number in the quick facts again. Nobody is reading 30,000 changes a day line by line, which is why the record has to exist.

You don't need any of this to check an AI's work.

When Claude searches the web or works through a task, you can expand its steps and see every search it ran and every page it opened.

Here's how I'd read them:

Retriever AI 

A free browser agent that finds leads, sends outreach and fills in job applications for you, with no API key or subscription.

Toby 

Mock interviews with an AI hiring manager. If you're a fresher, it's a low-stakes way to practise before the real thing.

Pancake 

Runs go-to-market experiments from Claude or Codex through one connection to 30+ data providers. Built for founders and marketers.

I think “can I see what it did?” is about to become the first question people ask before paying for any AI tool.

Right now agent products compete on how much they can do for you, and very few show how they got there.

That gets harder to accept once agents start touching money, client work and anything with your name on it.

A tool that can’t show its steps will start to feel like hiring someone who won’t show their work.

It also changes what being good with AI looks like.

Prompting got most of the attention over the last two years.

The skill I’d put time into now is reading an agent’s steps and spotting where it went wrong, and most teams haven’t started building it.

Hit reply and tell me if you’d pay for a tool that won’t show you its steps.

Until next time,
Vaibhav 🤝🏻

If you read till here, you might find this interesting

#AD 1

ChatGPT is insanely powerful.

But most people waste 90% of its potential by using it like Google.

These 1,000+ proven ChatGPT prompts fix that and help you work 10X faster.

Sign up for Superhuman AI and get:

Claim your free prompts

#AD 2

The key to maintaining brand quality at scale? Mastering how to communicate with AI models and embedding them into your creative strategy.

Join our honest discussion with Ari Murray, Chief Digital Officer at Salt and Stone, and get 5 tips for expert LLM prompting with ready to run prompts.

Register and get the prompts