The night shift nobody asked for

Anthropic named the feature "dreaming." That's not a metaphor they stumbled into. While you sleep, your Claude agent reviews its past sessions, spots recurring mistakes, and commits insights to long-term memory. The company announced it at Code with Claude alongside a batch of other capabilities, but the naming tells you where they think the value is: in the hours between your conversations.

Three announcements this week share a thread that most coverage treated separately. Dreaming lets Managed Agents learn from their own errors during idle time. Claude Code Routines let developers schedule agents that run on Anthropic's cloud against GitHub repos, no laptop open, no terminal session, just PRs waiting when you wake up. And Google DeepMind took a minority stake in Fenris Creations to test AI in EVE Online's player-driven economy, a game famous for rewarding exactly the kind of long-horizon planning that current AI can't do.

The conventional reading: AI is getting more capable. The more interesting reading: the unit of value is shifting from the session to the gap between sessions.

For most of the past three years, AI products have competed on how good they are while you're using them. Faster responses and better reasoning. The implicit economic model is attention: you sit down, you prompt, you get output, you leave. These three announcements invert that. Dreaming generates value from past sessions you've already closed. Routines generate value from tasks you haven't started yet. DeepMind's EVE Online research is explicitly about training agents that can sustain strategies across days and weeks, not seconds and minutes.

This pattern has a name in manufacturing: lights-out production. Factories that run without human operators on the floor, usually overnight. The economics are compelling. You've already paid for the building and the machines, so every hour they sit idle is waste. The same logic applies to AI infrastructure. Anthropic built the cloud capacity for Routines (4 vCPUs, 16 GB RAM per run). DeepMind negotiated access to isolated EVE Online servers. The fixed costs are sunk. The marginal cost of an agent working at 3 a.m. is nearly zero.

What this means for builders

If you're building products on top of AI, the strategic question is changing. It's no longer "how do I make the interaction better?" but "what can I have running when nobody's interacting at all?"

Anthropic's early numbers hint at the payoff: dreaming improved task success by up to 10 percentage points. That's not from a better model. It's from the same model given time to reflect. The compound effect matters here. An agent that fixes one recurring mistake per dream cycle gets measurably better each week without a single model update.

The EVE Online play is the most speculative of the three, but arguably the most revealing. DeepMind chose a game where the interesting decisions unfold over months, alliances shift, and markets respond to player behaviour. Current AI benchmarks reward fast, correct answers to bounded problems. EVE rewards patience, adaptation, and the willingness to lose a battle to win a war. If DeepMind cracks that, the applications reach well beyond gaming.

The companies making these bets are pricing in a future where AI's value accumulates like compound interest — slowly, in the background, while you're not looking. The question for everyone else: what's your product doing at 3 a.m.?

---