Persistent storage, real-time APIs, and constrained decoding models all shipped this week, moving agents closer to long-running production roles. The pattern is clear: infrastructure providers are building primitives that let agents work across sessions, process documents synchronously, and manage context without hitting rate limits or latency walls.
Introducing Vercel Drives: persistent storage for agents
Vercel Drives solves the stateless sandbox problem by giving agents persistent file storage that survives across executions. An agent can write to a drive in one sandbox, then mount that same drive in a completely different runtime with different tools installed. This matters for workflows where an agent needs to iteratively refine output, fork experimental branches, or hand off work to a specialized subprocess without serializing everything through a database or object store. Storage scales to 16 TiB and supports drive sharing and forking.
Vercel and Shopify make agents first-class shoppers
Vercel Shop is an open-source Next.js storefront template that treats AI agents as peers to human shoppers. It exposes machine-readable product catalogs and implements the Universal Commerce Protocol (UCP), letting agents browse inventory, compare items, and complete purchases programmatically. Shopify rewrote the Hydrogen SDK to work across JavaScript frameworks, so merchants can deploy agent-optimized storefronts at their own domains. This isn't just an API add-on, it's a signal that commerce platforms expect autonomous buyers.
Power real-time applications and agents with Document Intelligence
Databricks released synchronous REST APIs for document parsing, extraction, classification, and decision-making, replacing batch ETL pipelines with real-time calls. Agents can now invoke ai_parse_document or ai_extract mid-workflow and get structured data back immediately with citations. The use case here is interactive: a contract upload form that auto-fills key terms as the user waits, or an invoice that routes itself the moment it arrives. This turns document-heavy processes from overnight jobs into synchronous agent tasks.
const result = await databricks.aiExtract({
document: contractPDF,
schema: { parties: "string[]", termMonths: "number", penalties: "string" }
});
// Returns structured data with source citations in real timeHow Orange Business Built Live Intelligence on LangChain and LangSmith
Orange Business deployed an agent platform to 120,000 employees using LangChain for orchestration and LangSmith for evaluation, saving 12–18 months they would have spent building custom infra. Their 3-layer eval system covers RAG quality, agent behavior, and production trace monitoring, with centralized registries for agents and MCP servers to enable cross-team reuse. The takeaway: enterprises are treating agent deployment like application deployment, with CI/CD checks, trace-level debugging, and regression prevention built in before code ships.
Building Long-Running Agents with ScaleDown and Deep Agents
ScaleDown's constrained decoding models handle summarization, extraction, and compression tasks 200x faster and 467x cheaper than frontier LLMs. The langchain-scaledown package provides middleware for condensing message history, compressing tool outputs, and detecting context drift without paying frontier-model prices for every housekeeping operation. It's now supported in the LangSmith gateway with full tracing, meaning you can see exactly where context management happened and how much it cost.
These releases share a theme: agents need infrastructure that treats multi-step, multi-session work as a first-class concern. Persistent drives, real-time document APIs, and specialized context models are all answers to the same question—how do you keep agents running long enough to finish complex tasks without burning tokens or waiting on batch jobs?