Anthropic's own AI safety bill, a UK Microsoft probe and cheaper agents

The most interesting thing in today's news isn't a model release. It's a quiet admission that we have been building agents faster than we have learned to trust them. Amazon's new evaluation framework for Bedrock AgentCore addresses a problem anyone who has shipped an agent knows well: it looks brilliant in the demo, then drifts in production, and nobody notices until a customer does. Amazon is essentially selling the missing discipline, continuous evaluation, as a product. I think that is the right instinct, and it tells you where the real bottleneck now sits. It's no longer capability. It's confidence.

Anchor's Action Index attacks the same gap from the cost side. Instead of having a model reason through every web click from scratch, it stores proven workflows as shared execution memory and replays them as deterministic code. In a 300-task test, a small model using it matched a frontier model's success rate, with token costs reportedly down by a factor of 100. One benchmark isn't a verdict, and I would want to see how it holds up on messy, changing websites. But the idea matters beyond this one product: the most reliable agent may be the one that thinks the least, because it has already seen the task before. By the way, that is also a quiet argument for evaluation. Deterministic replays are far easier to test than free-form reasoning.

Trust is the theme in policy too. Politico reports that Anthropic has handed US lawmakers a draft AI safety bill of its own writing. I find this both sensible and slightly uncomfortable. Sensible, because legislators often lack the technical depth to draft workable rules. Uncomfortable, because a leading lab shaping its own regulatory framework is a textbook case of influence worth watching closely. Meanwhile, the UK competition regulator has proposed Strategic Market Status for Microsoft, putting Excel and Copilot under closer scrutiny. Taken together, the picture is clear: regulators on both sides of the Atlantic are moving from asking whether AI needs rules to deciding who gets to write them and which incumbents they bind.

Then there is the interpretability essay arguing that around 90 percent of neurons in neural networks serve a clear function. If that holds, it would change how we think about auditing models, and it would make the safety debate far less abstract. Researchers are split, and I understand the scepticism. Claims of easy readability have a habit of dissolving at scale. Still, it's worth keeping an eye on, because every safety proposal, whether written by a lab or a legislature, quietly assumes we will eventually be able to look inside these systems.

For builders, the practical takeaway is modest but real: GitHub's agentic autofix is now generally available, and EmbeddingGemma 2 brings multimodal search to edge devices, keeping data local. The tools are maturing. The open question is whether our habits of testing, auditing and oversight can mature at the same pace.