The month the chatbots grew (and threw) hands
For the best part of two years, our relationship with AI has lived inside a text box. You typed a prompt, watched a cursor blink, and received an essay, a recipe, or a block of Python. It was useful and occasionally brilliant, but it was always a very well-read intern locked in a glass booth, able to advise, and never able to touch anything.
In the last fortnight of May 2025, the booth acquired a door. Microsoft Build, Google I/O, a new Claude model family, and OpenAI's \$6.5 billion agreement to buy Jony Ive’s hardware start-up all landed within days of each other. The common thread was not a better chatbot; it was a chatbot with hands. If I hear the word "agentic" once more this week I may well unplug the router, but the shift underneath the vocabulary is real, and it deserves more scrutiny than the press releases are giving it.
Two Giants, One Pitch
Microsoft, and Google, arrived at the same conclusion at the same moment: nobody wants to babysit a chatbot any more. In Seattle, Build was dominated by the "open agentic web", a pitch about deploying specialised digital coworkers that talk to each other rather than helping you draft an email. The clearest evidence sits inside GitHub Copilot, which is graduating from autocomplete to something closer to an asynchronous peer programmer. You assign it a ticket, go and make a coffee, and return to a pull request waiting for review.
While Build was still running, Google took the stage at Shoreline and swung harder than it has in years. AI Mode is rolling out in Search across the US, Gemini 2.5 Pro gains a "Deep Think" capability that spends visible effort on awkward problems, and Project Mariner, the browser-driving agent, is moving from demo towards navigating ugly web forms. Then came Veo 3, with native, synchronised audio: footsteps on gravel, café chatter, lip-synced speech. The uncanny valley now looks more like a crack in the pavement, and we have stepped over it.
The Engine Behind the Hands
If the hyperscalers spent the week selling the vision, Anthropic spent Thursday 22nd May supplying a candidate for the engine. Claude Opus 4 and Sonnet 4 arrived to early reports of sustained focus across hours of tool calls, and terminal commands; with developers handing over large, tangled codebases, and walking away. These are promising signs, but an internal webinar is not a regression cycle, and we should treat the claims accordingly.
What has already changed is the argument. Six months ago the debate was whether AI had hit a scaling wall. Today it is whether senior developers are still writing syntax at all, or have been quietly promoted to managing a squad of tireless, slightly neurotic agents. The fear that this erodes craft is valid; but it confuses a change in where the skill sits with the loss of the skill itself. Nobody asked us, yet here we are...
A Hand Without a Screen
OpenAI, never keen to let a rival own a news cycle for more than a day, announced its purchase of io, the company founded by former Apple design chief Jony Ive. After the Humane AI Pin collapsed and the Rabbit R1 became a bright-orange paperweight, dedicated AI hardware looked like a fool's errand, and the smartphone appeared to have won for good.
Perhaps it has. But the designer of the iMac, the iPod, and the iPhone may know something about screen fatigue that the rest of us do not, and his remarks on our dependence on rectangles of glass suggest OpenAI is no longer building a better chatbot app. Add the reported move on Windsurf, and the cloud-based Codex agent, and the strategy is plain: own the model, the software, and the object in your pocket. It is Apple's playbook, applied to the 2030s.
Who Holds the Keys?
When an AI writes a bad poem, you laugh and close the tab. When it holds read-write access to your repository, your calendar, your browser session, and your internal databases, the risk profile changes overnight. Permissions, audit trails, and review capacity stop being footnotes and become the product.
The second-order effect is the one worth watching. Agents are superb at producing change, but humans review it at human speed. Every team that celebrates a surge in pull requests without asking who will read them is quietly building a queue that ends in either a bottleneck or a rubber stamp. Neither is an engineering strategy. Unless you DGAF about slop.
The Pragmatic Path Forward
The answer is not to refuse the hands; it is to be deliberate about what they are allowed to hold.
- Scope the permissions: Give an agent the access you would give a new contractor on day one, then widen it as trust is earned and evidenced.
- Make tests the specification: Treat every agent-authored change as untrusted until an automated suite says otherwise; the tests, not the author, are the authority.
- Budget for review: Measure capacity to verify change, not volume of change produced, and stop rewarding output for its own sake.
We spent two years learning how to talk to machines; the next 12 months will be about deciding what we let them touch.
Glorious, and terrifying.