Hi there,

AI agents are starting to run experiments, write software and make calls with little human input. The people building them are now asking how far that should go, and who gets to say when it has gone too far.

At The Information, we break the stories that matter before they become mainstream. Whether you’re deciding what to build, where to invest or how to compete, timing matters. Subscribe and save 25% on your first year to get this reporting as it happens.

Here are five stories on testing AI agents and setting their limits that we’ve reported in recent weeks.

How far should agents go

Sam Altman gave OpenAI staff three possible endgames for AI. He appears to back one and has a blunt word for another. Altman’s three endgames for AI explains why he is laying them out now.

The question is already playing out in a lab. Ginkgo Bioworks set out to pit OpenAI’s agents against a top scientist in what was billed as biology’s Kasparov-versus-Deep Blue moment. Shortly before the first round, the organizers pulled back. Inside the drama behind the contest, our Big Read, explains why.

Meanwhile, Anthropic is readying a model that one early user says picks up on what people really want without being told. The example that user gave shows how much initiative it is willing to take.

Who checks the work

OpenAI is in all-hands mode to get its next flagship model out safely, and not everyone agrees on how. We have the details on how OpenAI is scrambling.

And as AI writes more of the world’s software, companies say there is growing demand for a new kind of stress test that works like a flight simulator for software. We look at how it works.

The common thread: agents are being handed more to do while the question of who checks them is still open.

Before it’s everywhere, it’s in The Information. Subscribe for $299 and save 25% on your first year for reporting on AI, tech and finance.