Moonshot Releases Kimi K3.1 Pro, Promises Full Weights by Nov 4, 2026OpenAI Scales Habitat Storage for 1 Billion ChatGPT Users, 12x Faster ReadsAI Mindmeld Proposes Latent Interconnect Protocol to Bypass Text BottleneckRoleFlow Lets Multi-Role AI Agents Self-Evolve via Structural AdaptationMoEZip Compresses KV Cache by Routing to Cut MoE Memory UseMinimal-Edit On-Policy Distillation Fixes Small VLMs' Diagram ErrorsSelf-Assessment RL Teaches Search Agents to Know What They've FoundSWE-VERIFIER Cuts Coding Agent Costs with Selective VerificationVercel Explains How Its Agent Safely Acts on Production SystemsVercel CLI Adds Spend Management with vercel budgets CommandVercel MCP Adds submit_feedback Tool for Agent Bug ReportsVercel Connect Issues Restricted Stripe Keys per Agent or TaskVercel Deployments Now Show Build Cache Size and LimitsReplit Outlines Business Protections After Widespread Internet Vulnerability DisclosuresOpenAI Details Containment for Frontier Model Training EnvironmentsSimon Willison Tests Claude Painting Photos with Canvas in SecondsAI Is Getting Better at Math, and Researchers See a Predictable PatternMoonshot Releases Kimi K3.1 Pro, Promises Full Weights by Nov 4, 2026

In part two of its engineering series, OpenAI explains how it scaled Habitat with reliable multi-tenancy, global caching, and Azure Cosmos DB to deliver 12x faster reads for ChatGPT. The system serves online storage for more than 1 billion ChatGPT users, with multi-tenant isolation and caching strategy as core design choices.

A new arXiv paper, AI Mindmeld: A Latent Interconnect Protocol, addresses a bottleneck in LLM-based scientific agents. Models can consult specialists and run experiments, but specialist information must enter context as text, images, audio, or video, constraining transfer between models with rich high-dimensional representations. The protocol proposes passing latent representations directly between models instead of compressing information into text-like formats first.

A new arXiv paper introduces RoleFlow, a method that lets LLM-based multi-agent systems move beyond predefined collaboration structures. Existing systems typically rely on fixed role assignments and interaction flows, which limit adaptability as task requirements evolve. RoleFlow uses structural adaptation to let agents dynamically reorganize how they collaborate at runtime, optimized with trajectory-level rewards. The approach aims to improve adaptability to changing tasks. In practice, it could reduce manual work in designing multi-agent workflows and allow systems to reorganize roles on their own.

A new arXiv paper introduces MoEZip to address the memory bottleneck in long-context inference for sparse Mixture-of-Experts (MoE) models. MoE models keep compute in check by activating only a small subset of parameters per token, but KV cache grows linearly with sequence length, making memory the main constraint. MoEZip compresses the KV cache using routing information, leveraging expert activation patterns to decide which cached content is worth keeping. Practical impact: MoE models can support longer contexts with less memory, lowering deployment costs for long-text inference.

A new arXiv paper notes that visual math reasoning can fail before reasoning even starts: when a vision-language model misreads a diagram, all subsequent reasoning builds on false premises, and smaller models are hit hardest. The paper proposes minimal-edit on-policy distillation, whose core idea is to edit rather than imitate — the student model makes the smallest possible corrections to teacher outputs instead of copying them wholesale. Practical impact: better diagram-reading accuracy for small models on visual math problems, reducing reliance on much larger models.

A new arXiv paper targets search agents in retrieval-augmented LLMs. These agents can handle complex tasks by retrieving and synthesizing external information, but they often lack search-state awareness, meaning they can’t tell what information they already have and what’s still missing. The paper trains this ability with self-assessment reinforcement learning, so the model evaluates its own information state during search and adjusts subsequent retrieval accordingly. Practical impact: fewer wasted retrieval rounds and more reliable performance on multi-step QA and tasks that require repeated verification.

A new arXiv paper introduces SWE-VERIFIER to address wasted computation in coding agents. These agents spend significant compute generating actions and processing long interaction histories, and poor action choices push them into low-value states that raise API costs. SWE-VERIFIER samples and verifies alternative actions but verifies only selectively, aiming to preserve performance while cutting overhead. The paper positions it as a cost-efficient verification scheme. Practical impact: lower operating costs for coding agents, especially in long-running automated coding tasks.

Vercel published a blog post detailing how its agent handles permissions on production systems. The core problem: a request like “fix this deployment” doesn’t specify upfront which actions or access the agent will need. Broad access risks touching unrelated resources, while step-by-step approval slows users down. Vercel’s approach balances the two by scoping permissions to the task and adding approval steps for sensitive actions. The takeaway for engineering teams: before letting agents touch production, narrow permissions per task rather than granting account-wide access.

Vercel CLI now lets teams view and manage their Spend Management budget directly from the command line. Running vercel budgets returns current spend and the active budget, and users can also view, set, and remove budgets. It maps to the same On-Demand Budget shown in team billing settings. The change means developers, scripts, and agents can handle budget operations in the terminal instead of switching to the web dashboard.

Vercel MCP now includes submit_feedback, a tool that lets agents and harnesses send bug reports, friction reports, and feature requests directly to the Vercel team in plain language. The example given is asking an agent to report that a deployment error did not specify which environment failed. It lets agents file feedback at the point of discovery instead of requiring a human to relay it.

Vercel Connect can now create a separate restricted Stripe API key for each agent or task. When creating a Stripe connector, users can pick Managed API Key as the connection method; Vercel creates the key in the user’s Stripe account with permissions the user selects. The existing MCP and bring-your-own-key options remain available. Practical impact: less risk from sharing one Stripe key across agents, and errors from a single agent stay within its own permission scope, making per-task isolation easier.

Vercel now shows build cache size in deployment summaries, letting users see how much cache each deployment saved and how close it is to the limit for its build machine. The same data is available via the CLI using vercel inspect on version 65.1.0 or newer, and through the get-deployment REST API endpoint. Vercel notes that uploading the build cache is billed. Practical impact: teams can tell whether cache is nearing its limit before it affects build speed, and get a clearer view of cache-related costs.

Vulnerabilities disclosed this week affect software that much of the internet depends on, creating real risks for businesses and extra work for application security teams. Replit published guidance on how it protects business customers, noting that as AI makes vulnerabilities easier to find, teams will face growing pressure to remediate faster.

OpenAI published an explanation of how it strengthened containment for frontier model training and evaluation, including offline environments and high-compute red-teaming. Such disclosures typically address external scrutiny of safety boundaries around frontier training, with the focus on reducing exposure during the training process.

Simon Willison prompted Claude Opus 5.5 High to build an artifact that opens a photo and rapidly paints a version of it in a few seconds using a canvas element and a simulated paintbrush. He referenced an earlier experiment getting Claude Fable to simulate a watercolor illustration. The tests show how models can produce interactive visual artifacts that run in the browser.

AI models are solving more mathematical problems, and a pattern is emerging in their successes that could help predict which problems they will solve next. Rather than isolated breakthroughs, math capability appears to follow a traceable progression, which could inform how researchers benchmark models and choose test problems.