LangChain

An educational lab of AI agent architectures built on LangChain and local Ollama, offering multiple agent variants for chat, tool calling, RAG, hybrid, and agentic RAG modes.

OpenWiki Brains turns sources like Gmail, Notion, Git, X, Hacker News, and web search into a local wiki that agents can use as fresh, proactive memory.

Learn how to build a venture capital research agent that produces investment memos in 90 seconds with cited sources, using the Perplexity Agent API, LangGraph, and LangSmith. The agent runs parallel research nodes for team, financials, product, and market, then synthesizes a memo with seven sections, including a thesis and recommendation. Every claim is traced to primary sources, making the output auditable. The article also compares three search providers and offers takeaways for building similar agents.

A comparison of LangChain, LlamaIndex, and raw API calls for LLM applications, covering their strengths, trade-offs, and a decision framework for choosing the right abstraction level.

LangChain and NVIDIA launch the NemoClaw Deep Agents blueprint, combining Deep Agents Code, Nemotron 3 Ultra, and OpenShell for open, governed enterprise agents.

By tuning only the harness (scaffolding) around the Nemotron 3 Ultra, we achieved a best run of 0.86 on the Deep Agents suite, nearly matching Opus 4.8's best of 0.87, at roughly 10x lower cost. This article details the eval-driven approach, prompt engineering, middleware optimizations, and what didn't work.

NVIDIA Nemotron 3 Ultra is offering leading performance at lower cost than top closed models with the largest and most widely adopted AI agent orchestration platform. LangChain tuned its Deep Agents harness for NVIDIA Nemotron 3 Ultra, achieving the highest accuracy among open models, while completing more tasks at higher throughput and running at 10x lower inference cost per run than leading closed models.

Run Deep Agents Code on NVIDIA NemoClaw with deny-by-default networking, human approval, and audit logs for sensitive code modernization.

brAIn is an experimental AI agent framework that replaces the traditional chat loop with a NATS pub/sub bus architecture of long-lived daemon nodes. Nodes are reactive, only activating when relevant messages arrive, saving token consumption. Each node can have its own UI, supports distributed deployment, and features priority preemption and MCP client integration. The author demonstrates applications like ambient room agents, Slack listeners, and IoT controllers, and compares the architecture with existing tools such as LangGraph, AutoGen, and ROS 2.

This edition of AINews covers a broad range of AI developments from July 6-7, 2026. Highlights include Lilian Weng's deep dive into harness engineering for recursive self-improvement, Meta's launch of Muse Image and preview of Muse Video with agentic generation loops, and major product updates from Anthropic, LangChain, and Google on agent platforms. Other notable items: NVIDIA's Audex audio model, Cohere's Arabic ASR, robotics integrations with Hugging Face and NVIDIA, Liquid AI's Antidoom method to reduce reasoning loop failures, and Anthropic's controversial J-space interpretability work. Also covered: benchmarks for agents and legal AI, research automation, and inference efficiency advances.

How LangChain mines agent traces to find failures, fine-tune judge models cheaper than frontier LLMs, and hill-climb performance with evals.

Schneider Electric built enterprise LLMOps foundations with LangSmith to improve observability, evaluation, and deployment for AI products at scale. Their AI Hub of 350 experts deployed 60+ agents. The three pillars: self-hosted LangSmith for observability, offline/online evaluation with maturity framework, and per-product deployment. Case studies include internal AI assistant One Jo, CSM Copilot, and document processing agent, showing significant efficiency gains.

Deep Agents is an open-source agent harness by LangChain, designed for long-horizon, multi-step tasks. It includes built-in features such as sub-agents, filesystem access, context management, shell access, persistent memory, and human-in-the-loop approval. Model-agnostic and built on LangGraph, it is production-ready with LangSmith integration.

A detailed experiment comparing five Claude models (Opus, Fable, Sonnet, Sonnet 4.6, Haiku) on a full audit of the LangChain Python monorepo. Fable matched Opus in grade (A-) but excelled in generating actionable milestones and quick wins. The article presents findings, strengths/weaknesses, and recommends a multi-model pipeline.

Learn why coding agent bills spiral out of control — and how to trace, compare, and govern spend across Claude Code, Cursor, Copilot, and more in one place.

A comprehensive overview of 10 agentic AI frameworks in 2026, including LangGraph, CrewAI, OpenAI Agents SDK, Google ADK, PydanticAI, smolagents, Mastra, Microsoft Agent Framework, Strands Agents, and LlamaIndex Workflows, highlighting their strengths, best use cases, and trade-offs for developers.

OpenWiki generates and maintains codebase documentation so coding agents can find the repo context they need without loading everything into one instruction file.

Recursive Language Models (RLMs) combat context rot by having agents write code to dispatch subagents over context chunks. Deep Agents now supports RLMs through dynamic subagents and a lightweight code interpreter, enabling programmatic orchestration like map/reduce over large inputs. Benchmarked on OOLONG, RLMs outperform turn-by-turn agents at longer contexts.

Pendo used LangSmith to debug, evaluate, and monitor Novus, its AI product agent that turns behavioral data and session replays into code fixes. LangSmith's production tracing enabled Pendo to ship Novus in days with 90%+ success rate, saving 25% time on identifying new use cases and catching 60% of AI problems before customers noticed.

Evaluating long-running, stateful agents requires a new eval runner. Harbor integrates with LangChain's Deep Agents, LangSmith sandboxes, and observability to provide scalable, isolated evaluations with explainable traces.

The article introduces 'Wiki Memory' as a pattern for agent memory, where agents transform raw source data into a compact, persistent, and structured knowledge layer. Unlike RAG, it precomputes a higher-level synthesis. Examples include DeepWiki, Karpathy's LLM Wiki, and Factory's AutoWiki. The article also discusses open questions about raw data, compression format, and maintenance, concluding that files and agents are common answers.

This paper proposes an end-to-end agentic pipeline combining deep time-series forecasting, variational anomaly detection, and LLM reasoning to generate prioritized, actionable maintenance recommendations for office building appliance-level energy monitoring. The system uses a hybrid SSA-LSTM forecasting model and per-appliance LSTM VAE with attention for anomaly detection, with a three-stage LangChain pipeline (Context, Diagnosis, Report agents) featuring dynamic retrieval. Evaluated on a 16-scenario benchmark, the best backend scores 90.4/100 and a local 7B model passes all scenarios.

LangChain releases four new test environments to benchmark LLMs' ability to use tools effectively, covering function calling, planning, and reasoning. Tests include single-tool and multi-tool typewriter tasks, relational data queries, and a math task with altered rules. Key findings: GPT-4 excels on relational data but fails on longer trajectories; Claude 2.1 matches GPT-4 on three tasks; open-source models like Mistral 7b struggle with multi-step function composition; planning remains challenging for all models.

Compare GPT-4, Claude, and open-source LLMs on structured data extraction from chat logs. Benchmark results, evaluation metrics, and dataset creation insights.

Dynamic subagents let AI agents orchestrate work at scale using code instead of tool calls. Learn how programmatic orchestration in Deep Agents guarantees coverage, handles fan-out, and unlocks reliable multi-step, complex agent pipelines with common orchestration patterns and live traces.

Candidly built a state-aware conversational agent harness that uses an Input-Output Hidden Markov Model (IO-HMM) to infer user engagement states in real time from conversation traces, enabling targeted response policies that reduce disengagement. The system identifies four states—Engaged, Detailed, Guided, and Disengaging—and cuts disengaging turns from 23% to 11%.

Learn how Deep Agents uses prompt caching to cut LLM token costs by up to 80% across every major model provider - no extra config required.

New in LangSmith: a Fleet on-call copilot for alert triage, computer use for agents, voice trace debugging, and experiment status tracking. Plus Deep Agents Rubrics, programmatic subagents, a new LangSmith Deployment course, and upcoming events in Chicago, Berlin, DC, and Vegas.

On the Max Agency Podcast, Zack Reneau-Wedeen discusses the future of AI agents, advocating for simple architectures, outcome-based pricing, and avoiding 'org chart shipping.' He shares insights from building customer-facing agents at Sierra.

Klarna's AI assistant, built on LangGraph and LangSmith, handles the work of 700 full-time staff, reducing customer query resolution time by 80% and automating 70% of repetitive support tasks.

The EU AI Act compliance deadline is August 2, 2026. This article explains what the Act requires for high-risk AI systems and how LangSmith and LangChain OSS help meet each requirement through full observability, automated evaluations, human oversight, and more.

A practical guide to adding memory to AI agents, covering short-term and long-term memory concepts, trace analysis, and how LangSmith's tools enable a complete memory loop for agent improvement across runs.

This article explains how to build AI agents that can browse and interact with real websites using Playwright, browser-use, and LangGraph. It covers Playwright's advantages over Selenium (30-50% faster, persistent WebSocket, built-in auto-waiting, realistic events), setup steps, dynamic page scraping, multi-step form filling, anti-bot detection handling, session persistence, and Docker deployment. Through code examples, readers will create a working browser agent that navigates sites, fills forms, extracts structured data, and uses an LLM for decision-making.

LangSmith launches a no-code agent builder that enables non-technical users to create AI agents with memory, guided prompts, and MCP tools. The builder uses conversational guidance, built-in memory, and sub-agents to lower the barrier for agent development, suitable for internal productivity use cases.

This paper presents NAVI-Orbital, a software system on a LEO spacecraft that achieved the first in-orbit demonstration of a vision-language model performing autonomous multi-modal inference entirely onboard on April 16, 2026. Using Gemma 3 and LangGraph, it classifies scenes, generates descriptions, and responds to operator dialogue. Ground benchmark accuracy 88.16%, and it successfully processed uncorrected YAM-9 imagery onboard, demonstrating feasibility of semantic compression to reduce downlink bandwidth.

The author explains the motivation behind building a custom AI assistant instead of using existing tools, detailing the architecture, tech stack, and implementation process including LLM, LangChain, memory management, and tool integration.

This research addresses concurrency anomalies in multi-agent LLM systems by formalizing four anomaly types using TLA+ and building a mechanically verified consistency hierarchy L0-L4. With 274 Verus proof obligations, the detectors are proven sound and complete. Three Rust runtimes implement L0-L1, and the work reproduces real-world anomalies in ByteDance's deer-flow and LangGraph, providing verified fixes.

Factory AI leveraged LangSmith's observability and feedback API to close the product feedback loop, achieving a 2x improvement in iteration speed and significant reductions in development cycle time.

Open SWE is an open-source, cloud-hosted coding agent that autonomously handles GitHub tasks—planning, coding, testing, and opening PRs. It features a multi-agent architecture, human-in-the-loop control, and asynchronous execution.

Monte Carlo built an AI Troubleshooting Agent on LangGraph and debugged with LangSmith to help data teams resolve issues faster by exploring multiple investigation paths in parallel.

LangSmith launches public benchmarks and evaluation dataset sharing to help developers compare LLM architecture performance. The first benchmark is a Q&A dataset over LangChain docs, accompanied by the langchain-benchmarks package. The article analyzes various models and architectures, providing insights into performance and debugging.

LangSmith's homepage is now organized into Observability, Evaluation, and Prompt Engineering, with improved Resource Tags for flexible resource grouping. Onboarding guides and upcoming ABAC enhance usability.

Agent engineering is an emerging discipline that integrates product thinking, engineering, and data science to build reliable LLM agents through rapid iteration and production feedback. It addresses the unpredictability of agents by cycling through build, test, ship, observe, and refine, as practiced by companies like Clay, Vanta, LinkedIn, and Cloudflare.

Evaluate and compare fine-tuned open source LLMs using LangSmith. Test multiple models, automate evaluations, and choose the best performing AI.

AI agents work best when they reflect the knowledge and judgment your team has built over time. This article explores how to integrate human judgment into each stage of agent development, using a trader copilot example. It covers workflow design, tool design, and context engineering, and emphasizes the importance of automated evaluations and continuous iteration.

Learn how Deep Agents SDK manages context for long-running AI tasks through offloading, summarization, and filesystem abstraction to prevent context rot.

This post explores how to build reliable AI agents by designing loops, not just using a good model. It introduces four nested loops: the agent loop, verification loop, event-driven loop, and hill climbing loop, each building on the previous to create agents that work consistently and improve over time. Using LangChain primitives, developers can implement each level and embed human oversight where needed.

Fleet supports both quick, ad hoc tasks and recurring responsibilities. See how General Purpose Chat and Specialized Agents help teams delegate work.

This paper introduces a stateful ReAct agent built with LangGraph that avoids re-reading entire history at each iteration, cutting token costs by 90% in hyperparameter tuning and 52% in code optimization while maintaining performance. It provides a blueprint for practitioners to implement token-efficient autonomous experimentation. (Source: arXiv, June 2026)

LangChain and Fireworks fine-tuned an open model to mine perceived error signals from production traces, matching frontier model performance at a fraction of the cost.