Agents get serious, and the rules arrive
The EU's transparency rules land on Sunday, agents move from demo to work product, and the open-weights argument gets a formal position. What it means if you run a small UK firm.
- 01
EU AI Act transparency rules apply from 2 August
From Sunday the rest of the Act applies, with one carve-out: Article 6(1), the high-risk classification rule, waits until 2 August 2027. So the transparency duties land now. If a person is dealing with an AI they have to be told, and AI-generated or edited images, audio and video have to be labelled. It reaches UK firms whose AI output is used inside the EU, and nobody else.
EU AI Act implementation timeline - 02
Claude Opus 5 released, aimed at long-running agents
Anthropic shipped Opus 5 on 24 July, with the pitch aimed squarely at agents that run for a long time rather than answer one question. Coding and professional work are the named improvements. The direction of travel across the labs is the same: multi-step work, not single answers.
Anthropic - 03
Anthropic publishes a position on open-weights models
A formal statement on where the company stands on open-weight development and release, published 27 July. Worth reading if you care whether the tools you build on stay available on your own terms, which is the practical question behind the whole open-versus-closed argument.
Anthropic - 04
Frontier Red Team writes up three real security incidents
Published 30 July: an examination of three genuine cases from their cybersecurity evaluation work, used to sharpen how the evaluations are run. Notable because it is real incidents rather than benchmark scores, which is the kind of evidence a buyer can actually weigh.
Anthropic
The disclosure rule is the one owners resist, and it's the one that costs nothing.
An AI answering your phone should say it's an AI in the first sentence. The objection I hear is that callers will hang up. I don't think that's what happens. People don't mind a machine taking a message when the alternative is ringing out. They mind being tricked into thinking it was a person. Those are different things, and only one of them loses you the job. The firms that will struggle on Sunday are the ones whose whole pitch depended on the customer not noticing.
Worth knowing the edge of it: this reaches UK firms whose AI output is used inside the EU, and nobody else. If that isn't you, say it anyway, for the reason above.
The agent story underneath is the one that changes your week. Multi-step work means the useful question stops being "which AI tool should I buy" and becomes "which of my processes is messy enough to be worth handing over end to end". For most trade firms it's the same one: the call that lands at four on a Friday while everyone's on site, and nobody rings back until Monday.
I ran ventilation teams on UK sites. That call was never the emergency. Missing it was.
One process, done properly, with a human on the approve button. That beats eight half-connected apps every time.
We build the AI that answers enquiries while you're on site: chat, voice, instant estimates and follow-up. See how it works or price it in two minutes.