Klaviyo Ships 356 Internal Apps in Two Weeks on Vercel with 512 BuildersAgents Fail Consistency Test: 38%–74% Disagreement Across Repeated RunsIs Chain-of-Thought Decorative or Load-Bearing? Task Difficulty DecidesWhy Are Terminal-Bench Tasks Hard? Separating Real from Fake DifficultyRedundancy-Aware Learning Cuts Unnecessarily Long Reasoning TracesCOMED: Filling the Gap Between Routing and Collaboration in Multi-LLM InferencePrincipled Context Representation for Large-Scale Reasoning Across Vast SourcesGitHub Security Lab Launches AI-Powered Fuzzing Taskflow AgentVercel Opens Bug Bounty Program to Public After Years of Private TestingMeta Connect 2026: Muse Glasses, Voice, Video, and Charm UnveiledGoogle Launches Gemini 3.8 Live with Live Avatar FeatureClaude Code v2.1.282 Adds maxProseWidth Setting and Telemetry NoticesDatasette 1.0a41 Adds OpenTelemetry Support and Web Component Dialogscommit-rewriter 0.2 Adds Support for Non-Default BranchesOpenClaw 2026.9.6 Rebuilds macOS App to Fix Launch CrashGitHub Blog: When Chat Is the Wrong UI, Canvases Fill the GapFoundries vs Navigators: Why Thinking Got Cheap but Doing Didn't in ScienceOpenAI Codex Releases rust-v0.158.0-alpha.12Klaviyo Ships 356 Internal Apps in Two Weeks on Vercel with 512 Builders

Klaviyo built an internal app platform on Vercel that enabled 512 builders to ship 356 live apps in just two weeks. Every app ships SSO-gated and private by default, and builders go from idea to live app in about 3 minutes. Klaviyo is a B2C CRM serving more than 200,000 brands, spanning marketing, service, agents, and data. The platform dramatically shortens internal tool delivery cycles, letting non-specialist developers build and launch apps quickly.

A study ran 42 tasks three times each and found that, depending on the model, 38% to 74% of answers disagreed. The research notes that consistency is a baseline requirement for buyers, auditors, and regulators, yet current agents lack it. The paper argues skills should form habits to improve stability on repeat tasks. This directly limits agent usability in scenarios that demand reproducible, auditable results, such as compliance reviews and automated operations.

A study introduces continuation-based causal testing, an ablation-patch intervention that perturbs a single reasoning step, truncates the chain, and forces output, to test whether chain-of-thought monitoring is actually meaningful. The research finds CoT monitoring only matters when written reasoning causally constrains the answer, and that constraint depends on task difficulty. The takeaway: CoT monitoring may be decorative on easy tasks and genuinely load-bearing only on hard ones.

Frontier benchmarks need tasks current models cannot solve, but a task no model solves is not automatically hard. The research notes the same zero pass rate can come from a real capability gap, or from missing context or a broken reference implementation. The authors built an adjudicated agentic corpus to separate genuine hardness from fake hardness. This helps benchmark designers remove invalid hard tasks so evaluations more accurately reflect real model capability.

Large reasoning models often produce correct but unnecessarily long reasoning traces. Existing methods improve efficiency via trajectory-level objectives or local token- and step-level signals, but rarely model inter-step semantic dependencies. This work proposes a redundancy-aware learning approach that targets semantic redundancy between steps. The result is shorter reasoning traces while preserving answer correctness, which translates to lower inference cost and faster responses in deployment.

A new arXiv paper introduces COMED, a method for multi-LLM inference systems. Existing approaches either route among models or densely combine their outputs. However, routing stops after selecting an initial model, while dense collaboration is computationally expensive. COMED aims to fill the missing middle between routing and collaboration. For developers building multi-model inference systems, this method may offer a better balance between cost and performance.

Complex tasks in science, medicine, law, and finance often require assembling interdependent information scattered across vast, heterogeneous sources far beyond model context limits. Existing approaches attempt to address this, but the study proposes a principled context representation method. The core idea is to identify and retain what truly matters rather than simply stacking context. For professional scenarios requiring cross-document reasoning, this could improve accuracy and reduce interference from irrelevant context.

GitHub Security Lab published a blog post explaining how to use a new fuzzing taskflow built on its Taskflow Agent AI framework. The framework brings AI into security testing workflows, helping developers automatically find code vulnerabilities. The post details setup and usage for developers looking to integrate AI into security testing. The practical impact is that security teams can lower the barrier to fuzzing and locate potential vulnerabilities faster across their codebases.

Vercel has made its bug bounty program publicly available after running it privately through HackerOne since 2022. Over the past several years, Vercel worked with HackerOne’s VIP program to refine its process, targets, and scope, onboarding thousands of top researchers and hardening security across its platform. Those lessons have now been translated into a public program open to any security researcher. For developers, this means Vercel’s platform will receive broader community security scrutiny.

At Meta Connect 2026, Meta unveiled several new products including Muse smart glasses, along with new voice, video, and a feature called Charm. Latent Space described Zuckerberg’s team as being on fire. The announcement is primarily product-focused, with detailed specs and pricing not yet disclosed in the summary. The launch continues Meta’s push into AI hardware and multimodal interaction, and the combination of smart glasses with voice and video could reshape competition in consumer AI devices.

Google DeepMind released Gemini 3.8 Live, introducing a Live Avatar feature. Users can now interact with a virtual avatar in real-time conversations. The official blog post does not provide additional technical details or performance metrics at this time. For developers and users, this extends Gemini’s real-time interaction capabilities from text and voice to a visual avatar representation.

Claude Code released v2.1.282. A new maxProseWidth setting caps the width of Claude’s prose in wide terminals while tables and code blocks retain full width. The update also adds a startup notice and lists telemetry variables in a project’s settings files that were ignored or triggered, visible via /status and claude doctor. For developers using Claude Code in wide terminals, maxProseWidth improves long-text readability, while the telemetry notices help diagnose configuration issues.

Datasette released version 1.0a41. Alec Garcia added OpenTelemetry support in this release. Simon Willison refactored all of Datasette’s modal dialogs into a single Web Component, which is now documented for other plugins to use. Tags include JavaScript, Datasette, Web Components, Alex Garcia, and OpenTelemetry. For plugin developers, the new dialog component can be reused directly, lowering the cost of custom UI work, while OpenTelemetry support makes production observability easier to integrate.

Simon Willison released commit-rewriter 0.2. The key change is support for branches other than the default branch, allowing users to run the tool against another branch via uvx commit-rewriter –branch other. Previously limited to the default branch, the update addresses usage constraints in multi-branch development workflows. Developers working with multiple git branches can now rewrite commit history on any branch.

A GitHub Blog post explores UI design in developer tools, arguing that when developers need something more tangible and actionable than a chat box, canvases are the answer. The article discusses the limitations of chat interfaces in development scenarios, and how canvases provide a more intuitive interaction experience. For developer tool designers, it offers a framework for thinking about when to move beyond chat UI toward more structured interfaces.

A guest post on Latent Space argues that in science, thinking has become cheap while doing remains expensive. This asymmetry is quietly reshaping how research companies operate. The piece distinguishes between ‘foundries’ (execution-focused teams) and ‘navigators’ (exploration-focused teams), exploring how different types of research organizations are adjusting their positioning and operations amid this cost shift. For developers building AI-driven research tools, understanding this divide can help identify sharper product opportunities.

OpenAI Codex released rust-v0.158.0-alpha.12, following a series of alpha releases including 0.158.0-alpha.11, 0.157.0-alpha.11.1, and 0.158.0-alpha.10 through 0.158.0-alpha.7. The release summary does not include specific feature change descriptions at this time.