Software architectures for abundance
Software was designed around the assumption that intelligence is scarce and expensive. Most of its architecture — queues, forms, workflows — exists to ration human attention.
Scarcity shaped the stack
Nearly every abstraction in modern software exists to economize something scarce. Type systems ration human attention. Queues ration throughput. Forms ration data entry. Approval flows ration judgment. The stack is a long argument about how to use expensive human cognition carefully.
When intelligence becomes abundant and cheap, the premise inverts. The rationing machinery stops being an asset and starts being drag — dead weight that adds latency and indirection to systems that no longer need either.
What changes, what survives
Some abstractions survive abundance unchanged: databases, because durable state is orthogonal to intelligence; protocols, because agreement between independent parties is a coordination problem, not a cognitive one. Others are quietly dissolved: many intermediate formats, handoffs, and interfaces that existed purely to move work between specialized humans.
The open question is architectural. Systems where models are load-bearing components need different failure models than systems where software merely executes predetermined logic. What replaces the transaction when the actor inside it is stochastic? What does idempotency mean for a component that does not always do the same thing twice?
Research program
We are prototyping architectures where intelligence is treated as infrastructure — a utility with metering, fallbacks, and degradation modes — rather than as a feature bolted onto conventional software. The unit of design is not the model or the app but the system that composes them.