Search is no longer just ranked links. For agents, search is becoming programmable infrastructure.
As models gain control over retrieval pipelines, the hard problems become source quality, sandboxing, cost, repeatability, and auditability.
I see this as part of a broader shift from tools that answer to systems that act. Perplexity's Search as Code is interesting because it lets models assemble retrieval flows rather than only consume ranked results. That can improve task fit, but it also makes search harder to reason about operationally.
In regulated or high-stakes workflows, the question is no longer "did the agent find something." It is "which sources did it choose, under what constraints, and can we reproduce the path."
Programmable search will need the same discipline we expect from production workflows: logs, boundaries, and evidence.