LangChain describes a “knowledge factory” as a system that collects, curates, and maintains reusable company knowledge for agents, including skills, domain knowledge, and data-access guidance. The post recommends flagging knowledge gaps during agent work, using curator agents to maintain the information, and applying the approach to recurring workflows; LangChain says its internal GTM, content studio, and LangSmith Engine agents use variants of this approach.
This JavaScript/Node.js tutorial uses LangChain 1.6, LangGraph 1.7, and DeepAgents 1.16 to build agents from createAgent and Middleware through state management, tool-call context, retries, permissions, observability, virtual file systems, and sub-agents. Examples also cover Middleware testing, concurrent state updates, structured-output validation, framework selection, and code organization for operational systems. The guide notes that its examples use [email protected], @langchain/[email protected], and [email protected], and recommends checking updated API types after upgrades because the frameworks are evolving.
InfoQ 中文 AI&大模型 examines how agentic AI systems can be incorporated into microservices and cloud-native enterprise architectures through embedded agents, MCP tools, A2A connections, and asynchronous messaging. The article highlights context engineering, domain boundaries, task-specific orchestration, fine-grained permissions, and observability as core architectural concerns while arguing that microservices and cloud-native systems will remain foundational.
The report presents RAG as the mainstream approach to enterprise knowledge bases: clean and chunk documents, create both vector and BM25 indexes, retrieve relevant passages, and have the model answer from the retrieved text with citations. It recommends beginning with a limited scope and platforms such as Dify or RAGFlow before developing a custom pipeline. The workflow covers metadata filtering, permission checks before retrieval where document access differs, document updates and removal, and a fixed evaluation set. It also favors hybrid retrieval over vector search alone and warns that retrieval errors matter more than answer fluency.
Juejin’s final chapter of “Understanding RAG” reviews a 21-chapter course spanning embeddings, vector databases, retrieval, evaluation, production deployment, observability, and enterprise knowledge-base systems. It then surveys eight emerging directions, including Graph RAG, multimodal and agentic RAG, federated and on-device RAG, self-evolving knowledge bases, and RAG in the long-context era, before suggesting a continued-learning path.
Infinitus has introduced the Lingzhi Agent in its app to help frontline store staff find materials, draft sales scripts, coach teams and access business functions. According to a Juejin AI account of an Alibaba Cloud AgentCore salon, Infinitus uses AgentLoop and its Lingshu knowledge-operations system to process production traces, evaluate responses, capture human feedback and convert approved findings into reusable experience. The company’s engineers say the seven-stage pipeline links observability, data processing, evaluation, annotation, data return, experience mining and experience delivery, though its effectiveness has not been independently verified.
The project author open-sourced @yanisjiang/companion-core, a TypeScript runtime for adding state-driven companion behavior to existing applications. According to the author, traits and current state determine an intent, silence can bypass the LLM, and developers can connect the package to bots, game NPCs, desktop shells, or custom backends by supplying an LLM and storage. The repository’s Web/PWA implementation is presented as an optional example host rather than part of the core package.
A report argues that Python is more mature overall for large-model and agent development, with leading frameworks, a broader community and smoother tooling for vector databases, embeddings and streaming. Java offers strengths in enterprise integration, type safety, concurrency, observability and static analysis, but its frameworks remain fragmented and production guardrails often require custom implementation. It recommends Python for rapid experimentation and Java for controlled deployment within existing enterprise systems.
LangChain Blog says governed AI-agent deployments require centralized visibility and full run traces to detect costly, unauthorized, or unexpected behavior. It describes tracing as the foundation for verifying preventive guardrails and for production detective controls, while LangSmith Engine can analyze recurring failures and test proposed fixes. LangChain also reports that 87% of teams had tracing in place before first deploying one of more than 16,000 production agents on LangSmith.
A tutorial demonstrates how to build a dynamic few-shot prompting pipeline with LangChain. It loads a local bge-large-zh-v1.5 embedding model, uses Chroma and SemanticSimilarityExampleSelector to retrieve the three examples most semantically similar to each input, and combines them with the current review through FewShotPromptTemplate. The example constructs a sentiment-classification prompt but does not directly call a large language model.
LangChain announced three new LangSmith controls for managing access to sensitive production trace data: restricting trace inputs and outputs, assigning permissions through user groups, and defining trace-level policies from root run metadata. Administrators can let engineers retain access to run trees, latency, cost, and other debugging or performance data without exposing sensitive inputs and outputs; the controls are available in LangSmith Enterprise plans.
AWS Machine Learning Blog details three composable strategies used in Flightscape’s Training Schedule Optimization agent: offloading large artifacts to Amazon S3, paging tool results, and compacting context. The production agent handles JSON payloads larger than 20 MB and provides more than 40 tools for crew scheduling, resource allocation, and business-rule validation. On AWS’s internal benchmark with a 23 MB payload, the optimizations reduced token consumption by up to 71% and average latency by up to 80%.
LangChain shared what it has learned from running two internal proactive agents: GTM Agent, its sales agent that watches for new leads, sends daily briefs and drafts outbound emails, and Robin (formerly Open SWE), its engineering agent that drafts and reviews pull requests. The company says these agents act on three trigger types — real-world events, developer-set schedules and agent-set schedules — and that both were built on Managed Deep Agents, where more than 70% of runs now come from schedules rather than an ad hoc human prompt. LangChain argues that memory and review-focused UX are essential, outlining memory challenges such as salience, expiry and duplication, and a Slack-based interface where agents post work for approval and request reviews.
In the author’s implementation, one LangGraph Agent was connected through LangChain’s MCPAdapter to two independent FastMCP services. The yun120 service exposed 15 tools for local MySQL queries, statistics and charts, while demo1 provided flip_coin and make_decision. In multi-server stdio mode, MCPAdapter prefixed each tool with its server name, making tool names hard-coded in the system prompt unavailable. The author removed the default_ prefix from the yun120 tools, retained demo1 prefixes to avoid collisions and reported that the agent successfully called demo1_flip_coin. The planned production setup using multiple persistently hosted HTTP servers was not completed.
A commentary argues that AI agent design is moving from assigning discrete tasks to delegating outcomes and ongoing responsibilities. In this model, users specify a goal, while the agent decomposes it into tasks, updates its task list and iterates toward the goal. Systems need mechanisms for monitoring progress, obtaining human feedback, asking for clarification while running in the background, triggering work, and deciding when to stop, pause or take no action. Because required tool calls are less predictable, developers also need tighter controls over which tools an agent may use. Relevant decisions and feedback should be retained across runs.
The Canner/WrenAI repository describes WrenAI as an open-source generative business intelligence engine for AI agents, combining a semantic layer, an AI context layer, and governed text-to-SQL. Its Apache-2.0 engine includes an MCP server, CLI, and more than 20 connectors, while definitions are stored as reviewable YAML and Markdown in Git. The project also says Wren Engine merged into the main repository on May 7, 2026, while the earlier chat-first product was preserved as Wren GenBI Classic without new features or security fixes.
A Juejin AI article describes an unnamed internet platform's Agentic BI system, launched in early June 2026, with toolsets for semantic modeling, intelligent metric management, and Chat BI. The author reports more than 80% higher modeling efficiency in internal production validation, alongside additional gains in metric reuse and calculation speed, but provides no independent corroboration or measurement methodology. The account also documents failed generated queries, fiscal-year confusion, and safeguards using syntax and business-rule validation.
The Juejin AI report explains how to build a tool-calling agent in LangGraph by converting Python functions into tools with LangChain's @tool decorator and executing them through ToolNode. It shows how to bind multiple tools to a model, route requests using tool_calls, return tool results, and generate a final natural-language response. The article also compares the prebuilt ToolNode with a manual execution node and notes that invalid tool names or mismatched arguments require additional error handling.
The AG-UI project describes AG-UI as an open, lightweight, event-based protocol for connecting AI agents with user-facing applications. Its stated features include streaming chat, bidirectional state synchronization, generative UI, real-time context enrichment, frontend tool integration and human-in-the-loop collaboration. The project lists integrations spanning frameworks such as LangChain, CrewAI, Microsoft Agent Framework, Google ADK, AWS Strands Agents and LlamaIndex; OpenAI Agents SDK, AWS Bedrock Agents, Nim and Flowise are marked as in progress. It also provides a reference HTTP implementation, a default connector and a command for creating an AG-UI application.
The report explains contextual retrieval, in which an LLM generates a context description for each document chunk before the chunk is embedded for sparse and dense retrieval with reranking. It also describes Agentic RAG as an iterative process in which an agent decides when to search, what to search for, and whether to continue. These methods are combined in a two-layer memory design: a resident overview containing structured cards for key preferences, relationships, and summaries, and an on-demand details layer that retrieves original records and passages. The report attributes a qualitative improvement in offline BM25 recall@k to a book experiment using contextual prefixes, but provides no numerical results or independent reproduction.
A Juejin AI tutorial presents context engineering as a way to manage failures in long Agent conversations and tool-heavy workflows. It provides a roughly 60-line, dependency-free Python ContextManager that keeps system instructions resident, summarizes older history, preserves recent messages, and discards raw tool output after extracting key facts. The article also recommends budgeting summaries, repeating task-relevant rules near the generation point, and filtering tool results before they enter history.
LangChain argues that AI agents are moving beyond text generation, tool use, and decision-making toward pursuing open-ended goals, with Hone presented as an early example of this direction. It recommends defining optimization metrics and scoping each agent around goals that make high-quality human feedback easy to provide. In Actively AI’s case study, goal-oriented agents were reorganized from one agent per customer account into reusable sales “motions” built on Deep Agents, allowing context-gathering agents to assemble account information separately.
A Juejin tutorial proposes a layered memory architecture for stateless AI agents: in-process messages for one request, Redis for current-session context, and PostgreSQL or pgvector for cross-session memory. It provides Redis data-structure examples, a RedisMessageStore implementation using LangChain message mappings, an agent invocation loop, and production cautions covering TTLs, persistence, serialization, client configuration, and scaling.
LangChain announced a $300 million Series C funding round valuing the AI agent tooling company at $3 billion. CapitalG led the round, with GIC and DFJ Growth joining as new investors and IVP and Sapphire Ventures participating. LangChain says its LangSmith platform provides framework-agnostic tools for running, improving, and governing agents, and that more than 7,000 organizations use it.
An AWS post walks through a deployed reference architecture in which a LangChain-based enterprise orchestrator communicates with a Strands Agents-based SaaS agent across separate AWS accounts using the Agent-to-Agent (A2A) protocol. Three agents run in the two accounts with OAuth authentication in both directions. Each account has its own Amazon Cognito user pool, allowing each side to authenticate through its own identity provider without sharing secrets. The system is deployed through three Jupyter notebooks and tested end to end.
The Wall Street Journal reported that LangChain raised $300 million in a round led by Alphabet’s CapitalG, valuing the AI agent startup at $3 billion. The company helps large companies build and deploy their own AI agents and targets enterprises with an agent-development platform. The reported valuation is up from $1.25 billion in 2025; the financing has not been independently confirmed in the supplied material.
The MovieAgent project author published the second installment of a ten-part series, detailing a six-layer dependency model, a request path from HTTP input to streamed results, and three configuration switches for vector storage, embeddings, and LLM providers. The implementation uses a five-node LangGraph flow while reserving LLMs for intent parsing and response generation, with planning, execution, and fallback handling kept deterministic. On a self-reported sample of 1,134 films, Chroma, embedded Qdrant, and Milvus Lite showed sub-100-millisecond query latency, though the results have not been independently verified.
A Juejin AI tutorial describes an informal test in which the instruction “you are an interviewer” prompted DeepSeek-V3 to behave like a reassuring customer-service agent, even after an intentionally incorrect answer about thread pools. The tutorial replaces that one-line prompt with a four-part system persona covering role, interview rules, output boundaries, and language restrictions, then separates job profiles and candidate levels from reusable prompt templates. Its sample project includes configurable roles, prompt linting, and 43 offline tests.
LangChain says its internal GTM Agent generated 1,540 sales artifacts over 12 weeks, including account plans, proof-of-value documents, proposal decks and onboarding plans, saving representatives thousands of hours. The company attributes output quality mainly to business context, sales-stage logic and brand rules rather than prompt tuning, and says the same work can be replicated on the Deep Agents harness with Managed Deep Agents providing sandbox, skills and memory. LangChain also reports that deals where a rep sent one of these documents advanced to a later stage 47% of the time, versus 16% for matched deals without one, and that a stopping rule for self-review cut a one-pager run from 18 minutes to 7 minutes.
Google launched the Gemini Decisions API, powered by Gemini Decision 1, to evaluate up to 1,000 yes/no, multiple-choice, and rubric-scoring questions across text, images, audio, video, and PDFs in one pass, according to Google AI. The API is available in Google AI Studio, while Google reported 150–250 ms end-to-end latency for inputs up to 16K tokens; an independent developer test also found that it could force out-of-range inputs into existing categories when no fallback option was provided.
What problem it solves
AI agents often produce plausible SQL without the business definitions, approved joins, units, and examples needed to make results trustworthy. Teams also need a reviewable way to manage that context and turn governed answers into shareable dashboards across existing data sources.
Key features
Governed semantic and context layers — Business meaning is represented through the…
What problem it solves
Building AI chat experiences in React requires assembling chat primitives, runtime integrations, generative UI, and production UX details. The project provides a composable library that supports multiple backends, custom styling, and strongly typed APIs.
Key features
Composable Chat Primitives — Build a chat interface from primitives including Thread, Message, Composer, Thre…
What problem it solves
LangGraph provides infrastructure for long-running, stateful workflows and agents that need to persist through failures, incorporate human oversight, and maintain memory across sessions. It also supports debugging and deployment of agent systems with complex execution paths.
Key features
Durable execution — Build agents that persist through failures and run for extended peri…
What problem it solves
AI agents need structured access to crypto services spanning centralized and decentralized markets. The project provides reusable skills for tasks such as searching tokens, executing trades, tracking wallets, monitoring signals, and interacting with DeFi protocols through natural language.
Key features
Natural-language crypto access — Skills let AI agents search tokens, exec…