AI Daily · 2026-08-27

Qwen and Zhipu both launched new efficient open multimodal models today — Qwen3.8-Flash-Next and GLM-5.3-Flash — using hybrid sparse/linear attention architectures to cut training or inference costs to about a tenth of prior generations while targeting coding and agent workloads. Qwen activates only 6B of 125B parameters, supports a native 256K context that can scale to 1M tokens, and lowers production API pricing; GLM-5.3-Flash pairs 320B total parameters with 18B active, ships under MIT, and offers controllable reasoning effort. Google also introduced Gemini 3.5 Transcribe for multilingual low-WER speech-to-text and launched a double-blind evaluation pilot with confidential computing to reduce benchmark contamination while protecting model weights and evaluation data. In platforms and infrastructure, NVIDIA began shipping its agent-focused Vera CPU, LangChain announced a $125M raise alongside a wave of 1.0 releases, gateway and NVIDIA integrations, and Lovable argued SaaS is moving toward MCP-based agent interfaces. Other updates were more peripheral, spanning cloud gaming, earnings, EU AI Act compliance and glucose monitoring.

North America · First-hand

Google

⭐⭐⭐ [Research] Piloting the world's first double-blind AI evaluations

Google DeepMind is piloting the world's first double-blind evaluation of a proprietary frontier AI model, keeping external evaluation data in a cryptographic environment to prevent benchmark contamination. Working with the Singapore AI Safety Institute, OpenMined, AVERI, and MLCommons, the pilot tests a Gemini Flash Lite model against confidential benchmarks using Confidential Space within Google Cloud's Confidential Computing portfolio. In the double-blind setup, evaluators cannot see the Gemini model weights and Google cannot see the evaluators' test prompts, protecting both IP and evaluation data. Google says this strengthens trust in benchmark results, especially for highly sensitive evaluations such as cybersecurity or government use.

⭐⭐⭐ [Model Release] Intelligent transcription with Gemini 3.5 Transcribe

Google introduced Gemini 3.5 Transcribe, a new speech-to-text model for real-time intelligent voice interactions that turns raw audio into accurate, formatted text while handling background noise, jargon, and disfluency cleanup. The model already powers features in the Gemini app and on Android, such as Rambler on Android and the Gemini app on macOS, and is now available to developers via the Gemini API in Google AI Studio and the Gemini Enterprise Agent Platform, through two APIs: the Live API for streaming (gemini-3.5-transcribe-live) and the Interactions API for pre-recorded audio (gemini-3.5-transcribe). According to Artificial Analysis, it achieves an average WER of 4.0% for streaming and 2.6% for non-streaming use cases; it supports 85+ languages, custom vocabulary, speaker attribution for up to three speakers with word-level timestamps, and function calling to delegate tasks such as image generation and file analysis to other Gemini models. Compared with its predecessor Chirp 3, time to final transcription improves by 70%, and on the FLEURS benchmark it reaches 5.50% WER in streaming mode and 5.04% in non-streaming mode.

⭐⭐ [Research] GlucoFM: Foundation model for continuous glucose monitoring

Google Research released GlucoFM, a lightweight self-supervised foundation model for continuous glucose monitoring (CGM). It uses a dual-stream design that separately models slower glycemic trends and short-term deviations while preserving time-of-day and missingness. Pre-trained on 109,066 hours of unlabeled CGM data from Wear-CGM and four public datasets, it was evaluated on seven clinical prediction tasks across four cohorts (14 cohort-task evaluations). GlucoFM achieved an average PR-AUC 5.8 percentage points higher than the best GluFormer variant pre-trained on the same corpus, also posted the lowest mean absolute error for postprandial glucose response forecasting, and showed strong cross-dataset transfer and few-shot adaptation.

East Asia · First-hand

Qwen

⭐⭐⭐⭐ [Model Release] Qwen3.8-Flash-Next: A New Architecture, Towards Ultimate Cost-Efficiency

Qwen has open-sourced Qwen3.8-Flash-Next, a multimodal MoE model positioned as an early preview of the architecture behind the upcoming Qwen4 flagship, similar to the role Qwen3-Next played for Qwen3.5. The model upgrades attention, residual connections, embedding, and optimization: it uses a hybrid Gated DeltaNet + Qwen Sparse Attention design, introduces gated residual branches, N-gram embedding, and the Muon optimizer. The main model has 125B parameters with 6B activated per token, plus 51B N-gram embedding parameters; training cost is about 1/9 of Qwen3.7-Plus while delivering stronger coding and office-task performance. It natively supports 262,144 tokens of context, extendable to 1M tokens via YaRN. Weights are available on Hugging Face and ModelScope, and the production version Qwen3.8-Flash is priced at 0.16 USD per million input tokens and 0.47 USD per million output tokens on QwenCloud.

Zhipu GLM

⭐⭐⭐⭐ [Model Release] zai-org/GLM-5.3-Flash-BF16

Zhipu AI released GLM-5.3-Flash, the first natively multimodal model in the GLM-5 series. It has 320B total parameters with only 18B active, and the team says it outperforms GLM-5.2 across benchmarks and real-world workloads at one-tenth the price, approaching Claude Opus 4.8 on coding and agentic benchmarks. The model debuts a hybrid architecture combining sparse and linear attention in the GLM series, along with Manifold-Constrained Hyper-Connections (mHC), trained on a 30T-token multimodal corpus for greater compute efficiency. It can be served locally with SGLang, vLLM, TokenSpeed, Transformers, KTransformers, and Unsloth, and exposes reasoning_effort (low/high/max) and clear_thinking parameters to control the thinking budget. The model card is MIT-licensed and links to a blog and technical report.

⭐⭐⭐⭐ [Model Release] zai-org/GLM-5.3-Flash

Zhipu released GLM-5.3-Flash, the first natively multimodal model in the GLM-5 series, with 320B total and 18B active parameters. It introduces a hybrid architecture combining sparse and linear attention, along with Manifold-Constrained Hyper-Connections and a 30T-token multimodal pre-training corpus. According to the vendor, it outperforms GLM-5.2 across benchmarks and real-world workloads at one-tenth the price, approaching Claude Opus 4.8 on coding and agentic benchmarks. The model is open-sourced under the MIT license and deployable via SGLang, vLLM, Transformers, etc., with a reasoning_effort parameter (low/high/max) to control the thinking budget.

Ecosystem & Beyond (Products / Agents / Tools / Opinions)

Model Release

⭐⭐ [Model Release] Qwen3.8-Flash-Next

Qwen released a new open-weights model, Qwen3.8-Flash-Next, a multimodal MoE model described as an early preview of the architecture used in Qwen4. The model has 125B total parameters but only 6B active, which gives it a significant performance boost. Simon Willison tested Unsloth quantized versions on a DGX Spark, trying the 72.5GB UD-IQ1_S and 78.9GB UD-Q2_K_XL, and shared the generated images. He particularly liked the output from the UD-Q2_K_XL variant at high reasoning effort.

Product Update

⭐⭐⭐ [Product Update] Delivering Vera: NVIDIA’s First CPU Built for Agents Is Shipping Now

NVIDIA announced that Vera, its first CPU built for AI agents, is now shipping. The article says that as AI agents and trillion-parameter workloads become mainstream, new demands are placed on AI infrastructure, and its performance depends not only on compute (text truncated at this point). It also notes that NVLink Fusion is expanding with NVHBM custom high-bandwidth memory.

⭐⭐⭐ [Product Update] LangChain raises $125M to build the platform for agent engineering

LangChain announced a $125M funding round at a $1.25B valuation, led by IVP with participation from existing investors Sequoia, Benchmark, Amplify, and new investors CapitalG and Sapphire Ventures, to build the platform for agent engineering. The company also released LangChain and LangGraph 1.0, featuring a rewritten LangChain package focused on pre-built architectures for common agent patterns and improved model integrations, with no breaking changes until 2.0. New capabilities include Insights Agent in LangSmith and a no-code Agent Builder in private preview. LangSmith is positioned as a comprehensive platform covering observability, evaluation, deployment, and agent building. Customers cited include Replit, Clay, Harvey, Rippling, Cloudflare, Workday, and Cisco.

⭐⭐⭐ [Product Update] LangChain Announces Enterprise Agentic AI Platform Built with NVIDIA

LangChain announced a comprehensive integration with NVIDIA to deliver an enterprise-grade agentic AI development platform and is joining NVIDIA's Nemotron Coalition. The platform combines LangChain's LangSmith, LangGraph, Deep Agents and open-source frameworks with NVIDIA Agent Toolkit, including Nemotron models, NeMo Agent Toolkit, NIM microservices and Dynamo. It covers building, deploying, and monitoring production agents; NIM microservices offer up to 2.6x higher throughput than standard deployments, and Nemotron 3 Super's MoE architecture enables cost-efficient single-GPU deployment. The NVIDIA AI-Q Blueprint ranks #1 on deep research benchmarks, and LangSmith has processed over 15 billion traces and 100 trillion tokens.

⭐⭐⭐ [Product Update] Lovable CTO: The Future of SaaS Is Apps That Agents Can Use

Lovable CTO Fabian Hedin says the company is expanding from AI-powered app building into MCP-powered 'capabilities': selected functions in published apps are exposed as tools through a hosted MCP server, letting AI clients like ChatGPT and Claude call them directly, giving one app both a human UI and an agent interface. Lovable grew out of the GPT Engineer open-source project in 2023, went commercial and was rebranded in late 2024, and then saw users build production apps and internal tools on the platform. The company reports over $500 million annualized revenue run rate, 60M+ projects created, 900M+ monthly visits to Lovable-built apps, and usage by employees at nearly two-thirds of the Fortune 500. Menlo Ventures led a $400M Series C at a $13.3B valuation this month. Hedin attributes the fast pace to Lovable's innovation and rapid LLM improvements.

⭐⭐ [Product Update] GeForce NOW Gives Gamers More Ways to Play at Gamescom 2026

NVIDIA announced at Gamescom 2026 a series of GeForce NOW cloud gaming updates, including new DLSS 4.5 technology controls that let members fine-tune gameplay, plus expanded support for Steam devices, GOG single sign-on, and the Firefox browser, with more big PC games coming to the cloud. The company also highlighted new ways to play and broader device and platform support.

⭐⭐ [Product Update] LangSmith LLM Gateway: Runtime Controls for Agents

LangSmith LLM Gateway is now available in public beta. Operating as a centralized governance layer between agents and the models they call, it provides runtime controls including spend caps and rate limits at organization, workspace, API key, and user levels, plus model fallbacks across models and hosts. It also detects and redacts sensitive data such as PII and secrets before requests reach model providers. For multi-tenant setups, teams can use a custom request header to enforce per-customer spend and rate policies while routing all calls through one API key. The product is designed to let teams define policies once and enforce them consistently across models and agents, avoiding vendor lock-in.

⭐⭐ [Product Update] How LangSmith and LangChain OSS Help You Meet EU AI Act Requirements

LangChain's blog notes the EU AI Act compliance deadline of August 2, 2026, covering high-risk AI systems in finance, healthcare, HR, recruitment, biometric identification, and critical infrastructure, with penalties up to €15M or 3% of global annual turnover. The Act requires a risk management system, automatic event logging (Article 12), traceable decisions (Article 13), and data governance with bias review (Article 10). The post maps each requirement to LangSmith capabilities: end-to-end tracing of LLM calls and tool invocations, Studio visualization of execution graphs, Insights auto-clustering of failure modes, and custom dashboards with alerts. For retention, it offers self-hosted, BYOC, and EU-region managed cloud options, with base traces kept 14 days and extended traces 400 days. It also recommends continuous evaluation on production traffic for ongoing quality and safety scoring.

⭐⭐ [Product Update] Aligning LLM-as-a-Judge with Human Preferences

LangSmith introduces a self-improvement mechanism for LLM-as-a-Judge evaluators: human corrections are stored as few-shot examples and fed back into the evaluation prompts in future iterations. The goal is to make evaluators reflect user preferences more accurately without extra prompt engineering, adapting over time. The post explains why LLM-as-a-Judge has become popular (e.g., detecting RAG hallucinations/correctness and toxic outputs) and why it works: evaluators can use information not available at generation time, and judging is easier than generating. It also dives into implementation details and links to technical documentation and a video walkthrough.

⭐⭐ [Product Update] Announcing LangGraph v0.1 & LangGraph Cloud: Running agents at scale, reliably

LangChain announced the stable release of LangGraph v0.1. The framework, separate from the langchain package, is designed to give developers fine-grained control over agent workflows, including low-level control over code, prompts, and LLM calls, as well as conditional branching and looping for single-agent and multi-agent systems. The company also introduced LangGraph Cloud, currently in closed beta, as infrastructure for deploying LangGraph agents in a scalable, fault-tolerant way, along with an integrated developer experience for prototyping, debugging, and monitoring. The post says companies including Klarna, Replit, Ally, Elastic, and Norwegian Cruise Line are already using LangGraph, and includes customer quotes from NCL and Replit.

⭐⭐ [Product Update] How Podium optimized agent behavior and reduced engineering intervention by 90% with LangSmith

The LangChain blog highlights how customer Podium used LangSmith for dataset curation and finetuning throughout the lifecycle of their AI employee agent. As a result, the agent's F1 response quality improved to 98% and engineering intervention was reduced by 90%.

⭐⭐ [Product Update] Pushing LangSmith to new limits with Replit Agent's complex workflows

LangChain published a customer case study on how Replit and its Replit Agent pushed LangSmith to new limits. Replit serves over 30 million developers, and its Replit Agent went viral after launch, but the tool's complex agentic workflows generate very large traces involving hundreds of steps, straining data ingestion and display. To address this, LangChain and Replit worked together to add three capabilities to LangSmith: improved performance and scale for large traces, search-and-filter within traces to pinpoint issues, and a thread view that collates related traces from a single session to support human-in-the-loop workflows.

⭐⭐ [Product Update] Why We Rebuilt LangChain’s Chatbot and What We Learned

The LangChain team rebuilt their public chatbot chat.langchain.com after discovering their own support engineers weren't using it. Instead, engineers relied on a three-step workflow: reading official docs for intended behavior, checking the knowledge base for real-world issues, and verifying against the actual codebase. They automated this process with a Deep Agent composed of three specialized subagents — docs, knowledge base, and codebase search — whose results were synthesized by a main orchestrator agent with citations to documentation, support articles, and exact code line numbers. Engineers reported saving hours each week on complex debugging, so the team decided to apply the same architecture to the public chatbot, moving away from the old chunking-and-vector-search approach.

⭐⭐ [Product Update] LangChain and NVIDIA Launch NemoClaw Deep Agents Blueprint

LangChain and NVIDIA announced the NemoClaw Deep Agents blueprint for building enterprise agent systems, combining LangChain Deep Agents Code, the NVIDIA Nemotron 3 Ultra model, and the NVIDIA OpenShell runtime. The blueprint provides an open agent stack across model, harness, and runtime layers so teams can tune, secure, and govern their agents. On LangChain's agent eval suite, Nemotron 3 Ultra achieved an aggregate score of 0.86 at an inference cost of $4.48, versus $43.48 for the next-best model — roughly 10x lower cost. The core takeaway is that agent performance improves when the model, harness, evals, and runtime are tuned together, and that lower inference cost makes frequent evaluation and iteration practical.

⭐⭐ [Product Update] LangGraph Platform is now Generally Available: Deploy & manage long-running, stateful Agents

LangChain announced the GA of LangGraph Platform, purpose-built infrastructure and management layer for deploying and scaling long-running, stateful agents. Since beta last June, nearly 400 companies have used it to deploy agents to production. Features include 1-click deployment, 30 API endpoints, horizontal scaling, a persistence layer (memory, conversation history, human-in-the-loop/multi-agent collaboration), and the built-in LangGraph Studio debugging IDE; the enterprise tier supports RBAC. A note states that as of October 2025, the platform has been renamed LangSmith Deployment.

⭐⭐ [Product Update] Proving Agentic AI ROI in Financial Services | LangChain

LangChain and Pay-i jointly explain how to prove the ROI of agentic AI in financial services. They argue that multi-agent system costs are dynamic and multi-variable, driven by LLM calls, tool invocations, retries, and orchestration, which traditional FinOps tools cannot handle. The proposed solution combines LangSmith for per-trace observability and cost tracking with Pay-i to link those costs to measurable business KPIs. Two use cases are covered: RFP processing and AML compliance monitoring, with guidance on defining success metrics and monitoring them in real time. The post is a product-oriented engineering explainer without specific quantitative results.

⭐⭐ [Product Update] Build an Auditable VC Research Agent with Perplexity

The LangChain blog introduces an auditable VC research agent built with the Perplexity Agent API, LangGraph, and LangSmith. Given a company name, the agent produces a first-pass investment memo draft in about 90 seconds at roughly $0.40 in API cost, with every claim traced to a primary source. The architecture works in two stages: four research nodes (team, financials, product, market) gather evidence in parallel using Perplexity's web_search and finance_search tools, then a tool-less synthesizer assembles the memo into seven sections: Snapshot, Team, Financials, Product, Market, Risks, and Thesis. The post also covers LangGraph state management with a reducer that merges concurrent writes, and how LangSmith traces each line of the memo back to the search result that produced it.

⭐⭐ [Product Update] EU Macroeconomic Analysis with Deep Agents and You.com

LangChain published a partner blog demonstrating a macroeconomic research agent built with Deep Agents, LangSmith, and the You.com Finance Research API. The agent analyzes GDP data across all 27 EU member states, flags anomalous growth or contraction, and attributes causes to sector-level structural and cyclical drivers, producing a 13-section cited briefing in about 45 minutes. The You.com Finance Research API scores 87.29% on the FinSearchComp benchmark, and a full 27-country GDP run costs roughly $2.20 in API calls. In the demo, Ireland (12.3% GDP growth, pharma-led export surge with the industrial sector contributing +6.55pp alone) and Germany (structural contraction driven by automotive and construction, not a cyclical dip) were flagged as the largest outliers. LangSmith captures the complete execution trace so every data point in the final report can be traced back to the primary source that produced it.

⭐⭐ [Product Update] LangSmith Engine: How We Built an Agent for Improving Agents

This post describes the technical implementation of LangSmith Engine, which LangChain launched last week. Engine is an agent that sits on top of agent traces, identifies recurring failures, organizes them into actionable issues, and proposes durable improvements such as evaluators, dataset examples, and fixes. It takes two main inputs: an Agent Overview (a living instruction file similar to AGENTS.md) and traces from LangSmith tracing projects, and it updates its understanding of the target agent over time. Architecturally, Engine is an orchestrator agent that starts from compact trajectory summaries and loads full traces as needed to handle large trace volumes. Issues are shown in an Issue Board with fields including name, description, category, severity, evidence traces, proposed actions, and tags.

⭐⭐ [Product Update] Building a company due diligence Agent with Deep Agents

This recipe from the LangChain blog demonstrates how to build a company due-diligence agent that automates multi-step research using Deep Agents for orchestration and Parallel's Task API for web research. The agent coordinates five research tracks — corporate profile, financial health, litigation and regulatory, news and reputation, and competitive landscape — each handled by a dedicated subagent. Once the competitive landscape track returns named competitors, the orchestrator dispatches parallel competitor-analysis subagents, each running in its own isolated context. Parallel's Task API returns structured findings with per-field citations, reasoning traces, and confidence scores, while Deep Agents handles planning, delegation, and context management; the orchestrator cross-references findings and resolves discrepancies via Parallel's search API before writing the final report. The workflow was validated on Rivian Automotive with nine calls in about 23 minutes, and the post includes installation commands and tool definitions.

⭐⭐ [Product Update] Introducing Airbyte sources within LangChain

LangChain announced integration with Airbyte, introducing Airbyte document loaders that connect to 300+ data sources. Developers can load data directly in Python from platforms such as Stripe, Salesforce, and HubSpot for use in LangChain applications. The update targets developers building apps with external business data.

Research

⭐⭐ [Research] LangChain State of AI 2024 Report

LangChain published its 2024 State of AI report based on LangSmith usage data. OpenAI remains the most-used LLM provider, with usage 6x that of second-place Ollama; Ollama and Groq entered the top 5, and open-source providers collectively account for 20% of the top 20 providers by number of organizations. Chroma and FAISS remain the most popular vector stores, while Milvus, MongoDB and Elastic joined the top 10. 15.7% of LangSmith traces come from non-LangChain frameworks; the Python SDK accounts for 84.7% of usage, and JavaScript SDK usage reached 15.3%, up 3x year-over-year. Since LangGraph's release in March 2024, 43% of LangSmith organizations use it, and on average 21.9% of traces involve tool calls, reflecting the rise of agentic workflows.

⭐⭐ [Research] Evaluating OpenWiki with WikiBench

LangChain introduces WikiBench, a benchmark for evaluating their open-source documentation agent OpenWiki, designed to answer whether a wiki actually helps coding agents and whether OpenWiki improvements move the needle. Built on the Harbor framework, WikiBench pins a repository at a commit as the environment, has the agent generate a wiki, then uses a reader agent to answer coverage and retrieval questions about the repo. Answers are scored by LLM judges that check each expected fact and verify the fact is grounded in pages the agent read; a question with 3 of 5 facts scores 0.6. In an initial experiment comparing Bare DeepAgents, OpenWiki 0.2.5, and OpenWiki 0.3.0 with the same model (Luna), the ranking matched expectations, with OpenWiki 0.3.0 performing best.

⭐⭐ [Research] How Candidly Built State-Aware Agent Harnesses in LangSmith

Candidly built a state-aware agent harness for its AI financial planner Cait using LangSmith. They created a hybrid labeling pipeline that marked production conversations as resolved or abandoned using deterministic rules plus LLM-as-judge evaluators, and calibrated it against human-labeled data to reach 92.3% agreement. A gradient-boosted model trained on trace-derived features (e.g., Q/A alignment, topic continuity, user message length, caps ratio) separated resolved from abandoned conversations at 0.90 AUC. These features are cheap to compute in milliseconds at runtime, enabling Cait to infer user state mid-conversation and select response features for live steering.

Opinion

⭐⭐ [Opinion] Wiki Memory: File-Based Memory for AI Agents | LangChain

Harrison Chase proposes a wiki memory pattern for AI agents: an agent compresses raw sources such as logs, notes, code, and docs into a persistent, structured, inspectable knowledge layer, instead of retrieving raw chunks at query time like basic RAG. He cites existing examples including Cognition's DeepWiki, Factory's AutoWiki, and Karpathy's LLM Wiki, and distinguishes the approach from memory systems like LangMem, Letta, Mem0, and Zep by noting that wiki memory uses plain files as its substrate. He discusses open questions about raw data, compression format, and maintenance, with common answers being any agent-readable data, files, and agents handling both compression and updates. Files are attractive because they are inspectable, editable, and versionable. He concludes that wikis are best for durable domain knowledge and may be the simplest useful long-term memory pattern we have.

Other

⭐⭐ [Other] NVIDIA Announces Financial Results for Second Quarter Fiscal 2027

NVIDIA reported revenue of $96.2 billion for its fiscal second quarter ended July 26, 2026, up 18% quarter over quarter and 106% year over year. The announcement is a routine earnings release and provides no additional business or product details.

⭐⭐ [Other] How Factory used LangSmith to automate their feedback loop and improve iteration speed by 2x

LangChain published a customer case study: Factory, a company building an AI platform for SDLC automation, uses self-hosted LangSmith in its Code Droid for observability and feedback loops. Factory exports LangSmith traces to AWS CloudWatch to track data flow across LLM pipeline stages, and uses the Feedback API to attach positive/negative feedback to each LLM call, helping debug context issues like hallucinations. The feedback loop works as follows: Droids post comments in code reviews and collect feedback, LangSmith analyzes the data, engineers use custom LangChain tooling to optimize prompts and re-prompt the LLM, and feedback is exported to datasets for pattern analysis. According to the headline, this workflow improved iteration speed by 2x while reducing mental overhead and infrastructure requirements.

⭐⭐ [Other] Reflections on Three Years of Building LangChain

LangChain founder Harrison Chase reflects on the project's three-year journey, from a fall 2022 open-source Python package to a multi-language agent ecosystem with a commercial platform LangSmith, after ChatGPT accelerated adoption. The post announces a $125 million funding round at a $1.25 billion valuation to expand LangSmith and open-source efforts. He also shares his vision that LLMs connected to external data and APIs—agents—are powerful yet hard to build reliably, and the company aims to build the best tools to make reliable agents possible.

⭐⭐ [Other] AI Agent Latency 101: How do I speed up my AI agent?

LangChain co-founder Harrison Chase explains how to reduce AI agent latency. He advises first diagnosing the bottleneck, such as one large LLM call vs many small ones, and suggests using LangSmith's waterfall view to inspect per-step latency. To cut perceived latency, he recommends streaming intermediate outputs like plan steps, retrieval results, and thinking tokens, as well as running agents in the background. To reduce the number of LLM calls, he suggests combining code with LLM calls and using LangGraph to explicitly define how agents communicate. He also mentions using faster models and making LLM calls in parallel as common approaches.

⭐⭐ [Other] How Harmonic Rebuilt Scout on Deep Agents and 4x'd Retention

With LangSmith, Harmonic rebuilt its AI product Scout from a rigid architecture of composable subgraphs and per-node evals into a Deep Agents harness, where a single frontier model calls two tool sets: one querying a global data layer of 40M companies, 200M people, and 230K investors, and one accessing firm-specific context. The rebuild cut product iteration from months to days and quadrupled user retention. The UX shifted from a search box toward an investment-advisor style: users can hand over an investment thesis to get matching companies, or ask Scout to prep for a meeting from CRM records, emails, and LinkedIn data. The team's previous 80% good-outcome metric is no longer discussed, as consistent quality has become the default.

Later the same day

North America · First-hand

Anthropic

⭐⭐⭐ [Research] Previewing the Model Hardware Standard

Anthropic is opening a research preview of the Model Hardware Standard (MHS), a shared specification for AI agents to safely operate physical devices, initially with select scientific research labs and advanced manufacturers. MHS lets agents operate multiple lab and manufacturing instruments — such as microscopes, liquid handlers, and robotic arms — in parallel, for tasks ranging from routine drug discovery experiments to laser calibration on a quantum computer. Developed with HHMI Janelia Research Campus, MHS cuts hardware integration from weeks or months to hours or minutes via a standardized driver using simple "read"/"write" primitives and standard-format device discovery. The driver supports natural-language tags so agents can understand unfamiliar devices, works with any hardware that has a programmable interface, and is model-agnostic, accessible via MCP, CLI, or APIs. Anthropic plans to open-source MHS after collaborating with partners on safety evaluations and best practices for AI systems operating physical equipment.

⭐⭐ [Product Update] Expanding our support for scientists

Anthropic announced a significant expansion of its support for the scientific community. This follows the June launch of Claude Science, a product that integrates common research tools, produces auditable artifacts, and offers flexible access to computing resources. Starting today, Anthropic is opening 10,000 seats for scientists worldwide to access Claude subscriptions for one year, with standard seats free and premium seats with 5x usage limits priced at $15 per month, with plans to extend the program further. The AI for Science program is also broadening beyond biology to other fields, offering up to $50,000 in credits per project. Researchers in biology and chemistry remain limited to Opus-class models; Claude Fable models continue to block professional biology and drug development queries, and Anthropic is partnering with the US government on an access program for Mythos-class models in life sciences.

OpenAI

⭐⭐ [Research] Better answers, broader thinking: What students gain from ChatGPT and critical-thinking training

OpenAI published a randomized study of more than 1,000 students examining ChatGPT, critical thinking, originality, and student performance on a real-world university assignment.

Google

⭐⭐⭐ [Model Release] Gemini Omni 1.1 Flash lets you build with more control

Google released Gemini Omni 1.1 Flash, an update to its generative video model aimed at developers, available via the Gemini API in Google AI Studio and the Gemini Enterprise Agent Platform. The model now analyzes up to 10 seconds of prior video context (up from only the final second), enabling scene extension in 10-second increments up to a cumulative 40 seconds. Developers can specify start and end frames for smooth transitions, use 360p previews to iterate faster and cut costs, and upscale final output to 4K. Google describes the release as production-ready for professional use.

⭐⭐ [Product Update] 3 new ways to plan and book travel in Search

AI Mode in Google Search adds three travel-planning capabilities: flight price tracking with email alerts, available in over 180 countries and territories with data from more than 300 partner airlines and travel sites; points/miles cost display for flights and hotels, launching with Alaska Airlines, American Airlines, Choice Hotels, Hilton and Wyndham, with Accor, Flying Blue, Hyatt and others coming later; and direct hotel booking within AI Mode via Google Pay, with the hotel or booking platform acting as the merchant of record. Hotel booking is rolling out first in the U.S. in English.

⭐⭐ [Research] Planetary prediction engine: Automating global models via Earth AI

Google Research, under its Earth AI initiative, introduces the experimental planetary prediction engine (PPE), an autonomous system that handles the entire geospatial modeling workflow from data discovery to model training for tasks in public health, food security, environmental risk, and socioeconomics. Orchestrated by LLMs, PPE operates in three stages: intelligent geospatial data selection, multimodal dataset curation with a Feature Gate that mitigates target leakage, and automated model building and evaluation. It retrieves covariates from repositories such as Data Commons and Google Earth Engine, augmented by open-web discovery and pretrained embeddings like PDFM and AlphaEarth. The system reportedly reduces the time to build complex planetary prediction models from weeks of manual data engineering to minutes, while delivering improvements over diverse benchmarks; the paper is available.

Ecosystem & Beyond (Products / Agents / Tools / Opinions)

Product Update

⭐⭐ [Product Update] Start from scratch, without a repo

Cursor is updating Cloud Agents so they no longer require a connected GitHub or other third-party SCM provider. Users can pick「Start from scratch」in the repo picker and prompt the agent, with Cursor creating an Origin repo in the background. When satisfied, they can click「Create repo」to save the work as an Origin repo, choosing a custom name and private/internal visibility, and access it in the Codebase tab. Cursor also port-forwards the cloud agent's live environment to the browser for preview and design mode, and users can connect a Vercel account to publish their work to a live URL, with the publish feature requiring a Vercel account.

⭐⭐ [Product Update] Intelligent Model Routing on Replit

Replit announced the general availability of Intelligent Model Routing for all users. The feature automatically matches each evolving task with the most suitable model, balancing quality, speed, and cost behind the scenes, so users can focus on building rather than comparing models. No implementation details, model list, or performance data were disclosed.