The core shift is from asking one retrieval system for an answer to giving an agent bounded ways to explore, search, read, and verify its source corpus. The announcement is consequential because it connects retrieval behavior, tool exposure, and processing-cost attribution in one product surface.
New branch reason: The supplied related references contain no readable prior New Runtime record, while this source documents a distinct release bundle centered on retrieval control, MCP segmentation, and document-processing cost governance.
Thesis
LlamaIndex is turning document processing from a single retrieval call into a more modular, inspectable agent workflow.
Observed facts
Mechanism
Filesystem-style retrieval lets an agent enumerate files, use lexical search where appropriate, read source material, and combine that work with hybrid retrieval; task-specific MCP endpoints restrict the available tool surface to the operation at hand.
Why now
As document agents move into larger enterprise workflows, teams need both more reliable evidence access and a way to attribute processing spend by team, environment, customer, or product line.
Implications
Teams can design workflows that separate parsing, extraction, classification, retrieval, and verification rather than treating document intelligence as one opaque step.
Usage Tags, user metadata, and improved tier routing make cost a controllable dimension of the workflow rather than only an aggregate bill.
Evidence boundary
These are vendor-reported product capabilities and performance claims; the supplied source does not provide independent reliability results, comparative cost data, or production error rates.
Watch conditions
Watch whether LlamaIndex publishes evidence that narrower MCP surfaces and Retrieval Harness reduce retrieval or citation failures in real deployments.
Watch whether the distinct endpoints gain ecosystem support in automation and agent frameworks.