MiniMax introduces Mini Apps, Claude expands auto mode, and Agent Skills moves to a foundation
MiniMax is turning agent-generated pages into persistent Mini Apps, while Claude will make auto mode the default across its web, desktop and mobile apps. OpenAI's Decisions API reaches general availability with voice integration and larger batches, and Walmart is preparing separate authorization profiles and a developer gateway for shopping agents.
Elsewhere, the Agent Skills standard moves to the Agentic AI Foundation. AWS outlines local-and-cloud agent orchestration, and NVIDIA expands its enterprise data toolkit. New research examines formal software proofs, accurate voice data entry and the coverage of internal agent monitoring.
01
Products and Everyday Work
5 stories
2026-10-22MiniMax
MiniMax Code introduces persistent Mini Apps
MiniMax has introduced Mini Apps in local sessions of its Code desktop app for macOS and Windows. Generated applications can include a frontend, a Node.js backend and persistent data, remaining accessible in the conversation sidebar or a separate window. Official examples include a stock dashboard, a route planner and a CadQuery viewer that passes annotations on a 3D model back to the agent for editing.
The
/miniapp-creatorworkflow checks data sources with read-only queries, builds and publishes the app, then inspects it in a browser. The runtime manages processes, ports and publication rollbacks; manifests declare connectors and permissions, while the host holds credentials. App-generated messages enter the input box for review, and application code runs locally with the user's privileges.2026-10-22Anthropic
Claude will default to auto mode across web, desktop and mobile
Anthropic says sessions in Claude's web, desktop and mobile apps will default to auto mode next week for Pro, Max and Team plans. Instead of requesting approval at every step, an automated check blocks risky actions. Users can switch back to manual mode at any time.
This announcement extends the default to the everyday Claude apps; the earlier rollout applied to Claude Code. It specifies the eligible plans and clients but does not publish a new blocking-rate evaluation for these apps. Action screening remains part of the workflow, with automated checks replacing routine step-by-step confirmation.
2026-10-22OpenAI
Decisions API reaches general availability with 400 questions per request
OpenAI has made its GPT-6 Luna-powered Decisions API generally available, doubling the request limit from 200 to 400 questions. The API handles bounded decisions such as classification, scoring and choices, accepting text, images and native JSON. The update also permits images in question instructions and offers US and EU regional processing.
Native GPT-Live integration can trigger decisions as voice transcripts update, for example to identify a caller's intent and select the next workflow. Repeated-context caching lets successive requests reuse shared background information. OpenAI also reports improved latency in Asia. The additions extend the interface's voice integration, batching and reuse of common context.
2026-10-21Walmart
Walmart prepares authorization profiles and an MCP gateway for shopping agents
Walmart has announced Agent Profiles for Walmart and Sam's Club, with a launch planned for next month. Customers will authorize each agent separately, require approval for individual purchases or set limits such as $50, $100 or $200. Agent purchases will be separated from the customer's own activity, and permissions can be revoked or profiles deleted.
An accompanying MCP-based Agent Gateway is being built, starting with a beta for companies and developers. Walmart says agents will not receive payment-card details or be allowed to change payment methods; additional approval may still be required below a spending cap. Agents will see the same prices as human customers, and existing return policies will also cover mistaken agent purchases.
2026-10-21The Verge
Gemini arrives in Google Maps on CarPlay for conversational navigation
Google confirmed to The Verge that conversational Gemini in Google Maps is available to US CarPlay users, with a global rollout underway. Drivers can tap the Gemini icon in the car's Maps display, ask about places and follow up with navigation requests, adding preferences such as avoiding toll roads.
The integration also supports reporting traffic conditions through conversation, reducing the need to break a trip into separate searches. Specific regional rollout dates have not been announced. Google has not announced Personal Intelligence for CarPlay, including its connections to Gmail and Calendar; the confirmed capabilities concern conversation and travel actions within Maps.
02
Agent Tools and Open Standards
3 stories
2026-10-22Anthropic
Anthropic donates Agent Skills to the Agentic AI Foundation
Anthropic is donating Agent Skills to the Linux Foundation's Agentic AI Foundation, including the name, logo, GitHub organization and agentskills.io domain. Skills package task knowledge in folders containing a
SKILL.mdfile, scripts and reference material. Anthropic says more than 70 platforms have adopted the standard.Using or contributing Skills will not require foundation membership. Engineers from Anthropic, Google, OpenAI and other organizations will continue maintaining the project as individual contributors. Proposed next steps include versioned releases, a formal proposal process and security guidance. The transfer concerns governance and project assets; existing skills can continue to be used.
2026-10-21AWS
AWS demonstrates local routing and cloud experts with Strands Agents
AWS has published a Strands Agents reference implementation that uses a small local model to route tasks and sends complex analysis to Amazon Bedrock. The example runs
llama3.2:1bthrough Ollama as a router, bringing routine device-data queries, image-analysis tools and cloud expert analysis into one agent workflow.The tutorial uses industrial equipment monitoring to demonstrate telemetry retrieval, camera-image inspection and requests for expert insight. Its telemetry is simulated. This is adaptable sample code requiring Python, Ollama, Strands and configured AWS/Bedrock access; the post does not provide a production comparison of cost or accuracy.
2026-10-21NVIDIA
NVIDIA expands Agent Toolkit to connect agents with business data relationships
NVIDIA has expanded its enterprise Agent Toolkit with Agent Knowledge Foundation, bringing together CUDA-X data-processing libraries and capabilities including Auto-Ontology. That library maps raw enterprise data into business concepts and relationships, such as customers, products and orders, for agents to query and reason over. Kumo's structured-data models address tabular, relational and time-series analysis.
The toolkit also includes Nemotron Retriever for retrieval and reranking, and the OpenShell runtime to control agents' access and actions. NVIDIA describes integrations with Palantir, Databricks and Snowflake. Its own supply-chain example uses Palantir Foundry's business ontology and AIP to connect structured enterprise information to agent workflows.
03
Research and Reliability
2 stories
2026-10-22Anthropic
Anthropic and Theorem generate formal proofs for test entry points in 123 libraries
Anthropic and Theorem used Claude to generate roughly 180 million lines of Rocq proof code for 239 fuzzing entry points across 123 OSS-Fuzz libraries. After Claude fixes the bugs it finds, the project compiles targets to RISC-V and constructs mathematical proofs of crash freedom for inputs smaller than 64 MB. The Rocq kernel checks the proofs, extending beyond sampled test executions.
The conclusions apply to specified entry points, properties and an execution model: single-threaded static binaries, a fixed memory budget and defined RISC-V/Linux semantics. They do not establish that entire applications are free of every vulnerability. The team released a subset of the proofs and semantic definitions; confidence still depends on those definitions and the checking kernel corresponding to actual program behavior.
2026-10-22Sierra
Sierra's τ-Entity tests whether voice agents can capture details across three conditions
Sierra has introduced τ-Entity, a benchmark of 200 tasks across 10 entity types, including names, phone numbers, addresses and email addresses. Cooperative simulated callers will spell details when asked, and stored values must match exactly. Each task runs under regular, noise-heavy and speech-heavy conditions; Pass³ requires success in all three.
Across six languages, GPT-Live-1 scores 60.3% under regular conditions but 34.0% on Pass³. Separate experiments test a prescribed sequence of spelling, readback, confirmation and correction. Those configurations raise success across all three conditions from 14%–41% in self-directed conversations to 37%–54%, while adding roughly 21–28 seconds to average call duration. These protocol experiments use different configurations from the main leaderboard.
04
Security Research
2 stories
2026-10-22Redwood Research
Redwood and METR report on red-teaming OpenAI's internal agent monitoring
Researchers from Redwood Research and METR have published findings from an internal exercise conducted from July 21 to August 11, primarily examining OpenAI's monitoring system as it stood on or before July 21. The report discusses gaps in monitoring triggers after context compaction and checks that run only when control returns to the user. These mechanisms could leave some agent activity without timely inspection.
OpenAI staff had identified the context-compaction issue, which was fixed during the exercise. OpenAI also told the researchers that all discovered vulnerabilities affecting its production safeguards stack had been patched. The authors have not independently evaluated that broader remediation claim. The report describes the summer system and testing process, rather than measuring current production monitoring coverage.
2026-10-22Wiz
Wiz tracks ARTEX activity expanding to IT and software vendors
Wiz reports that ARTEX-related activity observed through October has targeted multiple IT and software vendors, focusing on identity systems, Microsoft 365 integrations and their APIs, including interfaces involving Okta, Entra and Cognito. Researchers connect the observed tooling and activity to Microsoft's Storm-0940 tracking designation through overlapping behaviors and tools.
Some observed automated sessions lasted 24–30 hours, switching IP addresses within minutes of being blocked and continuing to adjust requests. Wiz says most attempts were unsuccessful. The logs document persistent probing and adaptation; they do not establish that every targeted vendor suffered a compromise or data breach.