Agents Reshape Work While Quantum Leaps Forward

The way we build software is shifting under our feet. This week, the conversation crystallised around something that's been brewing for months: AI agents don't want sandboxes. They want actual computers. Fly.io's new Sprites-instant, durable, disposable Linux VMs that wake in seconds-represent something fundamental: we've been trying to fit agent workflows into containers designed for stateless, scalable applications. That friction is finally breaking.

Agents and the Machines They Actually Need

The pattern is becoming clearer. Anthropic's Claude Code, for all its capability, has hit friction with expensive rate limits. Meanwhile, Block's Goose-completely free, open-source, running locally or in the cloud-is gaining traction because it simply removes the constraints. More importantly, agents running on actual persistent machines can do things sandboxes can't: maintain development environments across sessions, access application logs in real-time, handle long-running processes that exceed arbitrary time budgets.

This matters because it's not just about convenience. Companies like Expedia are already asking how AI shapes their business strategy. Their CIO is experimenting with "agentic AI" to test new channels. Microsoft and Chubbies use Listen Labs to conduct customer research in hours instead of weeks. The infrastructure isn't the interesting part anymore-the question is what agents can actually do when you stop constraining them.

Building Better, Understanding Deeper

On the foundation layer, developers are getting better tools for understanding how systems actually work. The week brought substantive writing on circuit breakers, database transactions, and why sorting to check if something's sorted is wasteful. This isn't trendy-it's the unglamorous engineering work that separates products that scale from products that break. PlanetScale's explanation of how MySQL and Postgres handle concurrent writes differently shows what happens when you really understand your infrastructure. Fly.io's deep dive into their new orchestration stack demonstrates the same principle: decisions cascade, and architecture that works for one problem (scaling production apps) fails for another (instant, disposable development environments).

Quantum computing continues its incremental, genuinely exciting progress. Majorana qubits are now being demonstrated and manipulated. Iceberg Quantum's claims about reducing RSA-breaking qubits from millions to 100k-if they hold-would be remarkable, though the assumptions matter enormously. The field is moving from theoretical to engineering-focused: how do we actually build fault-tolerant systems, and what trade-offs exist?

Worth noticing: the best technical writing this week wasn't about flashy announcements. It was about fundamentals-transactions, checkpoints, error handling-written by people who've thought deeply about how systems fail and how to prevent it. That's where the real leverage lives for builders.