News

Agentic observability with Amazon OpenSearch Service MCP Apps

Amazon OpenSearch Service now supports MCP Apps. AI agents can return interactive visualizations beside their text responses inside a chat window. A single locally run MCP server now lets an agent move from an alert through traces and logs to a root cause in one conversation.

Evidence stays beside the diagnosis

Each response has a structured text summary and an interactive visualization. Available views include trace waterfalls, service topologies and log patterns. Charts and service maps are also supported. The setup connects a local MCP server, an IDE and an OpenSearch UI application. Users can inspect and verify each step inline without leaving the IDE. (AWS Machine Learning)

thinkidiot take: Moving from alert to root cause in one conversation changes how I would investigate a failure. I would keep the trace waterfall, service map and log patterns beside the agent's explanation, then verify its path inline. That is more useful than a chat response that asks me to trust an unseen tool call. Observability belongs inside the investigation, and this implementation puts the evidence where the claim is made.

What FID Hides: Detecting, Ranking, and Diagnosing Deviations in Generative Evaluation

An arXiv paper submitted on August 25, 2026 argues that FID and KID can conceal important differences between generated and reference distributions. On ImageNet, unrecognizable images tuned to match the reference Inception mean and covariance scored FID 24.7, while held-out real images scored 58.6. The paper introduces ZID to detect, rank and diagnose departures that a scalar FID gap does not calibrate against sampling variation.

A lower score can describe the worse images

FID summarizes only the first two moments of a distribution. ZID instead combines six standardized arms sensitive to location and dispersion. It uses a rank graph and Gaussian kernels at two bandwidths. Its outputs are a departure-magnitude index, a permutation p-value and a signed dispersion readout. Tests covered guidance sweeps for DiT-XL/2 and SiT-XL/2. ZID classified diversity collapse at high guidance as under-dispersion. (arXiv)

thinkidiot take: A FID of 24.7 for unrecognizable images, against 58.6 for held-out real images, is enough to stop treating the scalar as a verdict. I would pair the departure index with the permutation p-value and signed dispersion readout before ranking a run. That makes diversity collapse visible as under-dispersion instead of burying it in one score. An evaluation metric that rewards the wrong images deserves demotion, even when it is familiar.

Recursive Experiential-Working Memory Evolution for Long-Horizon Agent Harnesses

A paper submitted to arXiv on August 25, 2026 introduces Recuris, a memory architecture for agents working on long tasks. It improved success in 35 of 37 completed model-benchmark pairs across four benchmarks and ten models. On the longest tasks, its advantage reached 32.2 points, while common long-horizon failures fell by up to 80%.

Task state and learned skills stop competing for attention

Recuris separates Working Memory from Experiential Memory. Working Memory tracks task progress, while Experiential Memory guides skill selection. A fixed Meta-Agent makes localized updates to Skill Memory only after validation. On tau-bench, Recuris added 17.8 points to GPT-5.6 Sol. It added 15.6 points to Claude Opus 5, bringing that model to 87.9%. The design grounds skill use in the current task rather than a growing history alone. (arXiv)

thinkidiot take: Improvement in 35 of 37 completed pairs makes memory structure the result I would test first in a long-running harness. I would use Working Memory for current progress and let the validation gate control changes to Skill Memory. The 32.2-point advantage on the longest tasks shows how expensive an undifferentiated history becomes. Separating task state from experience is a better design than asking one expanding context to do both jobs.

5 ways to upgrade your home decor with Google Search

Google published five ways to use Search for home decorating and DIY work on August 25, 2026. The guide arrived after searches for “home decor inspo” rose 300% in the previous month. Users can now create furniture mockups from a room photo and dimensions, identify decor from a picture and get spoken DIY guidance with optional live video.

The room becomes the search query

AI Mode accepts a room photo and measurements such as an 84-inch-wide wall. It then generates furniture mockups for that space. Google Lens on iOS and Android identifies photographed decor and returns links to similar items online. Circle to Search finds similar products without requiring users to switch apps. It works on the latest Pixel and Samsung Galaxy S26 devices. Search Live answers spoken questions and can use live camera video to give step-by-step DIY instructions. (Google)

thinkidiot take: A 300% rise in “home decor inspo” searches explains why Google is turning the room itself into an input. I would upload a photo with the 84-inch wall measurement before comparing furniture, then use Lens for an item already in view. Keeping product discovery and DIY guidance tied to the actual space is more practical than searching from text alone. The strongest part is not the mockup, but the handoff from seeing an object to finding or working with it.

Granite 4.2 LLMs: How They're Built

IBM released Granite 4.2 as dense decoder-only models in 3B, 8B and 30B sizes. Each was pretrained from scratch on about 15 trillion tokens and finished with a 512K-token context window. All three support native tool calling and can emit OpenAI function-calling-format calls through an OpenAI-compatible endpoint.

Open weights arrive ready for tool-driven work

Training ran through five phases for every model. The 8B and 30B versions also received agentic reinforcement learning. That work covered tool use, code editing and execution, terminal operation and web search in sandboxed environments. Each model offers thinking, non-thinking and low-effort thinking modes. The architecture uses Grouped Query Attention with 40 attention heads and 8 KV heads. It also uses RoPE with θ = 10,000,000, SwiGLU and RMSNorm with ε = 1e-5. IBM released every Granite 4.2 model under the Apache 2.0 license. (Hugging Face)

thinkidiot take: Three Apache 2.0 models with a 512K-token window give me a real range for tool-calling experiments. I would start with the 8B model because it received the same agentic reinforcement learning categories as the 30B version while staying below its size. OpenAI-compatible function calls also reduce the integration work needed to compare them. Releasing the whole family with native tools is the right choice because agent behavior should be tested in execution, not inferred from model size.

Trending AI Papers

Ranking source: Hugging Face Papers for 2026-08-26.

WeMM-Embedding: WeChat Multi-Modal Embedding Technical Report

WeMM-Embedding is a family of models that places many kinds of content into one comparable representation. It handles writing, pictures, video, visual documents, and mixtures of these formats. The goal is to help systems find, sort, and recommend related material even when it comes in different forms.

• Problem: AI systems need a common way to compare content across several media types. Existing open-source models leave room for better accuracy, especially when inputs combine formats or require fine distinctions in relevance.
• New idea: The models create embeddings, which are compact numerical representations used to compare content. They accept text, images, videos, visual documents, and inputs that alternate between these forms. Training first teaches broad connections among media, then improves fine distinctions with selected examples, detailed relevance labels, and knowledge passed across model sizes. Users can also choose different output dimensions, meaning different lengths for the numerical representation.
• Simple example: It is like a librarian who can place a recipe, a cooking photo, and a demonstration video on the same shelf because they express the same idea.
• Evidence: The 2B model beats the previously leading 8B open-source baseline on MMEB-v2. The 9B model reaches a new best overall score of 80.6. The family also improves a 26-task internal benchmark and all 14 reported online A/B tests.
• Limitation: The abstract does not report detailed results for individual benchmarks, tasks, or deployed applications, so it leaves unclear where the gains are strongest or weakest.
• Why it matters: A shared representation for many media types can improve search, recommendations, classification, and agent systems.
• Paper: WeMM-Embedding: WeChat Multi-Modal Embedding Technical

AutoSaddler: Automatic Harness Optimization with Durable Updates from Agent Execution Traces

AutoSaddler is a system for improving the setup around an AI agent. That setup includes its instructions, available tools, and rules for deciding what to do next. The system studies failures and updates this setup automatically so agents can complete longer tasks more reliably.

• Problem: Agents often fail on long tasks because small mistakes accumulate as the work continues. External harnesses can reduce these failures, but people currently spend substantial effort testing prompts, tool settings, and control rules by hand.
• New idea: A harness is the collection of prompts, tools, and control logic that guides an agent. AutoSaddler treats improving that harness as an offline learning task, meaning it learns from saved execution records rather than changing the agent during a live task. It diagnoses batches of failure traces, proposes focused code-like patches, and validates the candidates before keeping an update. This process repeats so successful changes remain in the harness.
• Simple example: It is like a driving instructor reviewing several recorded failed trips, changing one part of the route guide, and testing that revision before using it for future drivers.
• Evidence: AutoSaddler improves the base harnesses by 9.0 percentage points on GAIA2, 9.6 points on SWE-Bench Pro, and 10.0 points on Terminal-Bench 2.0. Ablation studies indicate that deep debugging, targeted changes, and selection for broader usefulness each contribute to effective optimization.
• Limitation: The abstract reports results on three benchmarks but does not test whether the improvements persist across other agents, tasks, or real-world deployments.
• Why it matters: Automating harness improvement could make long-running agents more dependable without requiring as much manual tuning.
• Paper: AutoSaddler: Automatic Harness Optimization with Durable

On-Policy Self-Distillation in Diffusion Models

DiffusionOPSD is a training method for making image generators better follow rewards such as human preferences or task goals. These generators build an image through a sequence of denoising steps. The method turns a final judgment about an image into specific teaching signals for predictions made along the way.

• Problem: A reward can score the finished image without explaining how any intermediate denoising prediction should change. This makes it hard to connect an end result to useful corrections during the generation process.
• New idea: On-policy means the training examples come from the model behavior currently being used. Self-distillation means one version of the model supplies structured teaching targets for another version. A frozen behavior policy, which is a temporarily fixed model, produces generation paths and reference predictions. Reward gradients then create bounded better and worse targets around those references, the trainable model fits them for a limited period, and an exponential moving average, a gradual weighted update, refreshes the behavior policy.
• Simple example: It is like a drawing teacher who turns a final score for a finished picture into small corrections at several stages of the sketch, then updates the lesson after seeing the student's latest work.
• Evidence: Across SD 3.5-M and Z-Image-Turbo, the method earns the best final held-out scores in 19 of 20 reward-matched settings spanning two backbones and ten evaluators. It beats the strongest competing method by up to 44.0%. Compared with DiffusionNFT, it cuts training GPU-hours by 40% on SD 3.5-M and 63% on Z-Image-Turbo.
• Limitation: The experiments cover two model backbones and reward-matched evaluator settings, but the abstract does not establish whether the method generalizes to other diffusion models, rewards, or human evaluations.
• Why it matters: Clear intermediate teaching signals can make diffusion model alignment more efficient and easier to analyze.
• Paper: On-Policy Self-Distillation in Diffusion Models

Trending AI Repositories

Ranking source: GitHub Trending.

anthropics/claude-plugins-community

Anthropic maintains this repository as the public record of community plugins that passed its review pipeline. Its nightly sync makes it useful for seeing what is available to install.

• What it is: This Python repository provides a read-only mirror of community plugins approved through Anthropic’s review pipeline. The marketplace list is stored in .claude-plugin/marketplace.json.
• What it does: Community plugin marketplace for Claude Cowork and Claude Code. Read-only mirror , submit plugins at clau.de/plugin-directory-submission.
• Who it helps: It helps Claude Cowork and Claude Code users find community plugins. They can inspect the marketplace list and choose plugins to install.
• Limitation: The repository is read-only, so plugin submissions must go through Anthropic’s submission page.
• Repository: anthropics/claude-plugins-community

TauricResearch/TradingAgents

This research project explores how several language-model agents can work together on financial trading. It is worth attention as an implementation linked to an arXiv paper and a trading research community.

• What it is: This Python research framework explores how multiple language-model agents can work together on financial trading and is accompanied by an arXiv paper.
• What it does: TradingAgents: Multi-Agents LLM Financial Trading Framework
• Who it helps: It helps researchers study agent-based approaches to financial trading. They can examine the framework, read the paper, and join the linked research community.
• Limitation: Understanding the project may require reading the linked arXiv paper.
• Repository: TauricResearch/TradingAgents

AgriciDaniel/claude-obsidian

This project builds an Obsidian knowledge base that improves as more material enters it. It stands out by keeping notes in files the user owns while supporting connected retrieval and maintenance.

• What it is: This Python project uses Claude Code to build and maintain an Obsidian vault of owned, connected notes.
• What it does: Self-organizing AI second brain for Obsidian + Claude Code. Drop any source and Claude reads, links, and files it into one connected knowledge graph of plain Markdown you own. AI note-taking, personal knowledge management (PKM), and an open-source Notion alternative. Based on Karpathy's LLM Wiki pattern.
• Who it helps: It helps people who want a personal knowledge base without giving up control of their files. They can capture sources, create connected notes, retrieve grounded answers, and keep the vault healthy.
• Limitation: It requires Obsidian and Claude Code.
• Repository: AgriciDaniel/claude-obsidian