Agentic Search: Alibaba Cloud Shifts AI Search from Answers to Task Execution

AI Search Enters Scale Phase: Value Shifts from Traffic to Task Outcomes

Generative AI adoption has accustomed users to expressing complex intent in natural language and expecting synthesized results. Public data shows growing user bases and usage frequency for Google AI Overviews, Google AI Mode, ChatGPT, and Perplexity, indicating AI search has moved from proof-of-concept to scaled application. A cited customer study found AI-search-sourced traffic conversion rates up to 9× higher than traditional search, though the figure varies by industry and attribution; the trend underscores a shift from raw traffic volume to intent quality and task outcomes.

From "I Know" to "Task Done": Enterprise Search Agents Move Toward Controllable Execution

Xing Shaomin outlined three stages of AI search evolution: pre-2022 keyword retrieval and link return; 2023–2024 question understanding, information synthesis, and answer generation; 2025–2026 task execution where users provide a goal and the system handles step decomposition, tool selection, knowledge access, operation execution, and result delivery. The endpoint shifts from "I know" to "task completed," requiring a system that connects knowledge, planning, execution, memory, and feedback—not merely a large model bolted onto a search box.

Alibaba Cloud Agentic Search builds on enterprise knowledge for scenarios like industry research, ops diagnosis, automated inspection, data insight, and e-commerce guidance. Its enterprise-grade Search Agent comprises Agent Loop, sandbox, tool calling, permission audit, Memory, AutoSkill, PageIndex, and AgenticWiki. It ingests multimodal documents, online collaboration docs, knowledge bases, search engines, and databases, using hybrid retrieval and dynamic orchestration to turn scattered data into citable, verifiable execution evidence.

In long tasks, memory must be accurately selected, updated, and forgotten—not simply maximized. Internal evaluations in September 2026 showed the memory module achieved 96.69% comprehensive score on LoCoMo and 81.37% overall accuracy on LongMemEval-V2 , while reducing input tokens by 68.18% under the LoCoMo test configuration. These results illustrate the system's balancing of effectiveness, cost, and context length.

Self-evolution is not unconstrained auto-modification. Tasks are completed by collaborating Planner, Researcher, Executor, and Reviewer agents. Execution trajectories, quality scores, error corrections, user feedback, and citation evidence feed a unified experience stream that drives Memory updates, Skill extraction and evaluation, and incremental knowledge-base deposition. Evolved capabilities are loaded on demand and remain observable, evaluable, and rollback-capable, preventing accidental experiences from hardening into system behavior.

Elasticsearch Agent Extends Search to Ops Diagnosis and Data Insight

The presentation demonstrated Alibaba Cloud Elasticsearch Agent for operations diagnosis and data insight. In ops scenarios, it aggregates cluster state, logs, metrics, and historical events to produce structured diagnostic reports containing anomalous nodes, key metrics, probable causes, suggested actions, and evidence chains. In data insight, users pose business analysis tasks in natural language; the Agent understands metric definitions, retrieves data, identifies issues, and forms improvement recommendations.

The product follows a "probe–analysis–insight" architecture: continuous collection of logs, metrics, events, monitoring, and alerts at the base; Agentic Search provides knowledge, memory, Skills, sandbox, and permission capabilities in the middle; and Agents, Tools, and Skills form intelligent inspection, fault diagnosis, performance tuning, and data insight at the top. Behind the simple natural-language entry point lie permission isolation, tool governance, execution audit, and evidence traceability as prerequisites.

From Retrieval System to Context Infrastructure: Elasticsearch Redesigned for Agents

When search directly serves agents, results must be structured, machine-readable, citable real-time context—not just human-readable snippets. Agents call more frequently, demand fresher data, involve more tenants, and require full-text search, vector search, multi-tenancy, stable latency, and controllable cost. Xing stated that search engines are becoming cognitive and context infrastructure for agents .

To meet these needs, Alibaba Cloud Elasticsearch built a multi-tenant engine for agents: write and query nodes are decoupled, using OSS as unified persistent storage; tenant data is organized by slices; cold data does not occupy compute or local disk long-term, loading into memory or SSD cache only on query. In a million-document test, hot-query and cold-query P99 latencies were approximately 60 ms and 600 ms respectively , and overall cost dropped about 70% under the described architecture and measurement scope (actual results vary with data scale, hardware, and access patterns).

Qoder serves as a flagship case for this agent engine. AI coding faces large code repositories and document corpora, many tenants, continuous code growth, and the need for real-time retrieval after updates. The disclosed Qoder scenario supports hundreds of billions of documents, tens of millions of code repositories, and hundreds of billions of vectors , with hot-query P99 around 60 ms and second-level code update visibility ; under the corresponding load and benchmark criteria, total cost is roughly one-seventh of comparable solutions .

A core enabler is the cloud-native kernel FalconSeek . It maintains compatibility with Elasticsearch APIs and Query DSL, moves core query execution to a C++ kernel that reads Lucene segment files directly, and reduces CPU, memory, and GC overhead through batched execution, memory lifecycle management, and query-specific optimizations. For vector search, FalconSeek first traverses the HNSW graph to obtain candidates, then reads raw float vectors for batched reranking, allowing full-text and vector capabilities to be co-optimized within a single engine .

For agent-generated long-term memory, skills, and knowledge, the Elasticsearch knowledge engine uses Namespace as the tenant and agent identity boundary . Within a unified index it fuses BM25, vector, filter, rerank, and personalized retrieval, and manages data of differing value and access frequency through Hot, Warm, Cold, and Frozen lifecycle tiers. This forms a "retrieval–action–memory–knowledge" enterprise knowledge self-evolution closed loop .

Xing concluded that Memory makes context accumulable, Skills make experience reusable, and the knowledge engine keeps enterprise assets flowing within security boundaries. Through the synergy of Agentic Search, Elasticsearch, FalconSeek, and the knowledge engine, Alibaba Cloud aims to provide complete infrastructure from search to action , delivering real-time, reliable, verifiable, and cost-controlled context support.

Source: 2026 Yunqi Conference AI Search Forum keynote by Xing Shaomin, Alibaba Cloud Intelligence Computing Platform Division AI Search Lead. All benchmark, performance, and cost figures reflect the specific versions, configurations, and test methodologies disclosed in the presentation.