Elasticsearch 9.4 sends a clear signal: search is no longer a "tool for humans," but Agent's "memory and action driver." From 9.0 to 9.4 in less than a year, four iterations have all centered on Agent infrastructure. What's worth noting: the discourse power of next-generation enterprise data infrastructure sits not in large models, but in the middle layer that can handle Agent concurrent calls.

What this is

Elasticsearch is upgrading from "full-text search database" to "Agent-era enterprise data foundation." Core changes come in three layers: vector search shifts from pure-memory to disk-clustered DiskBBQ, achieving 7x the throughput of Qdrant at equivalent recall; through native CPU SIMD instructions (CPU hardware-level parallel computing) and NVIDIA cuVS GPU acceleration, index construction speeds up 12x; the Agent Builder, Workflows, and Agent Skills trio launches, supporting MCP (standard protocol for Agent tool invocation) and A2A (Agent-to-Agent conversation protocol), feeding search results directly into the Agent reasoning chain. ES|QL stuffs retrieval, computation, and correlation analysis into a single pipeline, letting Agents use one language to complete the full "vector search + time filter + log correlation + LLM invocation" workflow.

Industry view

Positive voices argue that the real bottleneck for AI Agent deployment isn't the model, but whether enterprise data can be called frequently, cheaply, and accurately. DiskBBQ and GPU acceleration solve the real problem of "memory and compute costs under billions of vectors," with code already merged into Apache Lucene and a complete open-source path.

But the editorial team also sees several risks. First, this is a technical share authored by Elastic's China Chief Architect — essentially a product pitch, not an independent evaluation. All performance comparisons (vs Qdrant, Milvus) come from official benchmarks. Second, Elasticsearch's traditional strength is search and logging, but "Agent infrastructure" is a new front; competitors also include Databricks, ClickHouse, and other players embedding vector capabilities into data warehouses. Whether this is "full-stack integration" or "doing everything but nothing deep" remains to be seen. Third, the Agent orchestration layer's maturity is far behind open-source frameworks like LangChain and LlamaIndex; before enterprises pay up, they'll ask "why not use open source?"

Impact on regular people

For enterprise IT: over the next two years, AI deployment budgets will shift from "buying LLM APIs" to "rebuilding the data foundation." Vector search and Agent orchestration will become new procurement categories, and enterprises already deployed on ES will benefit first.

For individual careers: search and operations engineer skill stacks are being rewritten — understanding vector search, quantization algorithms, and Agent protocols is now required. Mid-layer talent who can articulate "how enterprise data serves Agents" will command a premium.

For the consumer market: no short-term awareness, but reduced enterprise IT costs will transmit to middle-platform, customer service, and other scenarios, ultimately reflected in whether the customer service rep on the other end of your phone call "gets you" better.