The future of code, science and autonomous agents

Today’s most important AI news

Today was a day of conversations about the near future: what happens when AI writes most of the code, when scientific models stop being support tools and become hypothesis generators, and when autonomous agents start training themselves. There was no single dominant launch, but an accumulation of signals pointing in the same direction: AI is ceasing to assist and starting to act.

Anthropic and the uncomfortable question about code: who’s programming here?

At its “Code with Claude” developer event held this week in London, Anthropic presented a vision of software development that made more than one attendee uncomfortable. The proposal wasn’t just that Claude helps write code, but that the entire workflow — from idea to deployment — can be orchestrated around the model. The developers in the room were asked whether they had recently “shipped” code, and the conversation quickly turned to how much of that code had been generated by the model versus the human.

This matters for any product or engineering team because it redefines the developer’s role. It’s not about whether Claude can write a working endpoint — it already can. The real question is what part of technical judgment, architecture and the decision of what to build remains the human’s responsibility. For now, that boundary exists, but Anthropic is actively pushing to move it.

The practical implication is concrete: if your team is still debating whether “using AI to program” is cheating or a competitive advantage, that debate is over. The industry settled it without waiting for you.

Google DeepMind bets that AI can do science, not just assist it

At this week’s Google I/O, Demis Hassabis, CEO of Google DeepMind, stated that we are “standing in the foothills of the singularity”. Beyond the rhetoric, what was shown was substantial: a reconfiguration of how DeepMind conceives AI’s role in science. The shift isn’t minor. Until recently, the narrative was that models accelerated existing research. Now the bet is that they can identify new questions, generate hypotheses and design experiments.

For a product manager or executive in sectors such as biotech, materials or pharma, this changes the investment analysis for AI tools. It’s no longer about operational efficiency — cutting hours of literature review or automating reports — but about generative capability: can this system find something no human would have looked for?

The risk is that the promise outruns the available evidence. Current models still make systematic errors in basic quantitative reasoning (benchmarks show accuracy rates of just 50-53% on elementary physics tasks), which makes it necessary to maintain expert supervision in any serious scientific pipeline.

Agents that fine-tune themselves: FT-Dojo and the end of manual fine-tuning

A paper published this week introduces FT-Dojo, a benchmark environment designed to evaluate whether language agents can autonomously handle the fine-tuning process for other models. The task includes data curation, training configuration and iterative diagnosis of model behavior — everything that today consumes weeks of specialized work in vertical AI projects.

The context is relevant: adapting an LLM to a specific domain — health, legal, manufacturing — is still expensive, slow and dependent on scarce expertise. FT-Dojo asks whether that process can become an agent task, where the system manages its own improvement iterations with minimal supervision. Complementing this, another work published today proposes directly “compiling” agentic workflows into the model’s weights, eliminating the need for external orchestrators and reducing operating costs by two orders of magnitude compared to current architectures such as LangGraph or CrewAI.

For teams evaluating AI deployments in specialized domains, this has build vs. buy implications: if fine-tuning can be automated and workflows can be “baked” into the model, the barrier to entry for proprietary models drops significantly. It’s not for tomorrow, but it isn’t distant science fiction either.

In a nutshell

What ties today’s news together is a shift in control. Anthropic wants the developer to supervise, not write. DeepMind wants the scientist to validate, not generate hypotheses. FT-Dojo wants the ML engineer to review, not configure. The pattern is consistent: AI absorbing the execution layers and leaving the human in the judgment layers. The problem is that “judgment” is hard to define, hard to measure and very easy to over-delegate without noticing. That’s the real tension of 2026, and none of these systems resolve it — they only make it more urgent.

---

Sources used: MIT Technology Review - The Download (https://www.technologyreview.com/2026/05/22/1137845/the-download-coding-future-steroid-olympics-ai-science/), MIT Technology Review - Google I/O and science (https://www.technologyreview.com/2026/05/22/1137813/google-i-o-showed-how-the-path-for-ai-science-is-shifting/), ArXiv - FT-Dojo (https://arxiv.org/abs/2603.01712), ArXiv - Compiling Agentic Workflows (https://arxiv.org/abs/2605.22502), ArXiv - Agentic Physical AI (https://arxiv.org/abs/2512.23292)