AI tools, models and workflows
The AI stories worth reading from the past 7 days: model releases, developer tools, agents and AI safety. The latest issue comes first, earlier days follow.
44 stories in the past 7 days
Latest issue8 stories
UK Regulators Publish AI Cyber Risk and Resilience Guidance
The UK PRA, FCA and Bank of England jointly published practical AI-related cyber risk and resilience guidance on November 19, supplementing existing cyber risk supervisory requirements. The document says AI can strengthen defenses but also brings new risks, and human oversight remains irreplaceable; highly autonomous, externally accessible AI systems carry higher risk and require proportionate controls based on an institution's size, complexity and risk profile. Recommendations include assessing critical business dependence on AI, reviewing agent permissions and external interactions, and avoiding over-reliance on AI detection. The guidance is not a new supervisory standard but AI-specific supplementary consideration.
Eclipse Enclave 1.0 Launches: Coding Agents Run Autonomously in Isolated Sandboxes
The Eclipse Foundation released open-source, vendor-neutral Eclipse Enclave 1.0, providing an isolated, policy-controlled execution environment for AI coding agents to operate autonomously. It supports agents including Codex, Claude Code and OpenCode in an extensible architecture, using isolated containers, controlled file and network access, API key protection and persistent sessions so agents can modify code, run commands, install dependencies and run tests without gaining unrestricted system permissions. Enclave 1.0 is MIT-licensed and available at enclave.eclipse.dev.
Google Launches Semantic Governance Policies to Review Agent Intent Before Tool Calls
Google previewed semantic governance policies on the Gemini platform, using Agent Gateway to check before an agent executes a tool call whether each proposed action complies with user-defined rules and original intent; violations are replaced with a rejection message and the decision is logged to Cloud Logging. Features include no-code policy authoring in natural language, baseline protection by default, support for multiple interfaces including the Gemini API, Claude via Gemini, OpenAI-compatible APIs and self-hosted models, and auditable decision records.
Google Data Agent Kit Generally Available; Coding Agents Operate BigQuery Directly
Google Data Agent Kit is generally available, giving external coding agents such as Claude Code, Codex and Gemini CLI native access to Google Data Cloud via the Model Context Protocol, covering BigQuery and more than 15 services. Agents can inspect schemas, run queries, read logs and manage resources, with skills for BigQuery SQL optimization, Bigtable key design and dbt pipelines. The tools are free, billed only for underlying cloud usage; agents can run as the user or an impersonated service account, inherit existing IAM and VPC Service Controls, and be governed via Model Armor.
Perplexity Brain Auto-Generates and Updates Personal Skills for Computer
Perplexity's Brain feature in Computer can automatically create and update personal skills—reusable rule-based workflows—based on users' past sessions, reducing manual setup. An eight-week production test showed 72.9% of users with automatic skill setup added at least one personal skill, versus only 6.5% in the control group. The system uses a four-stage loop of evidence, proposal, check and activation, validating personal skills with deterministic checks while built-in skills require broader testing and human approval; overall tool-call failure rates fell 11%. The feature is on by default for US individual and enterprise subscribers and can be disabled in settings.
Gemini API Launches Tool Search, Loading Large Tool Libraries on Demand to Cut Latency
Google previewed tool search in the Gemini API's Interactions API, letting agents avoid loading all tool schemas at once. After developers add the tool_search tool and set defer_loading for specified tools, the model first sees only tool names and descriptions, with actual schemas loaded within the same request when needed; search runs server-side with no self-hosting or index building required. The feature also supports remote MCP servers, with deferred loading pulling in all their tools on demand. Tool search itself is free, but loaded schemas are billed as input; it supports Gemini 3.6+ and all Gemini 4 models.
AWS DevOps Agent Adds Chat Capabilities via Slack, Teams and MCP Clients
AWS released a new version of DevOps Agent, expanding chat capabilities to MCP clients such as Kiro and Claude Code, as well as Slack and Microsoft Teams. Through the Model Context Protocol, the agent can perform instructed chat actions and integrate tools including GitHub, GitLab, Azure DevOps, ServiceNow and PagerDuty to automate workflows from chat interfaces. New features also include automatic Slack program channels for each incident, two-way Teams chat, a compute sandbox that can run scripts with network access, and cross-platform conversation renaming and archiving.
Amazon Textract Insights Generally Available: Verifiable Generative Document Extraction
Amazon Textract Insights is generally available, a new generative AI feature for Textract. Users define extraction needs with blueprints in JSON Schema form, can describe fields in natural language, and get support for verbatim reads and inferred values, along with per-field confidence scores and page-level evidence so results are auditable. The system handles mixed document packets of up to 3,000 pages, automatically splitting them and matching the corresponding blueprint, with provenance information including source pages and coordinates, and requires no model training. Users can start from public blueprints for W-2s, Form 1040s and passports, or customize their own.
5 stories
Firecrawl Launches Alexandria to Connect AI Agents to 550+ Specialized Data Sources
Firecrawl launched Alexandria, a hosted directory connecting AI agents to more than 550 specialized data sources including financial data, public records, scientific data, software documentation, and business contacts, sourcing content through web retrieval, dedicated indexes, and external providers. It can be used via Firecrawl's hosted service, API, or Model Context Protocol integration, with content composed of commercial sources and public data provided under vendor terms; Alexandria itself is not open source. The workflow has three steps: explore tools, review inputs/outputs and pricing, then execute; developers can call it via JavaScript/Python SDK, HTTP, CLI, or MCP. Agents must use the exchange parameter to select enabled providers and can pin...
MathWorks Connects AI Agents to Engineering Tools via MCP and Skills
MathWorks announced a new agentic AI framework connecting MATLAB and Simulink to AI agents through the Model Context Protocol and dedicated Skill packages. It introduced the MATLAB Agentic Toolkit, Simulink Agentic Toolkit, and MATLAB MCP Server, enabling AI coding assistants to execute, inspect, and iterate in real engineering environments. MCP exposes installed toolboxes, code execution, and unit tests, while Skill packages provide task-oriented guidance on which models, functions, and parameters to use and in what order. MathWorks emphasizes that human verification remains essential, managing risk through workflow controls and permissions. This lets agents...
OpenAI Reveals Performance Factory: Using Codex Agents to Automatically Hunt Performance Regressions
OpenAI explained how its Performance Factory uses Codex agents to automate ChatGPT performance engineering. The system automatically identifies, investigates, fixes, and verifies performance regressions, centered on a reusable workflow: engineers define tasks and prompts, tasks advance by state, a scheduling loop automatically dispatches ready work to Codex agents, and execution runs in the cloud via internal APIs, with a subscription mechanism waking agents on events like deployments or comments. A built-in perf-forensics skill guides Codex to find the onset of performance declines, affected requests, and possible deployment or configuration changes. The first regression workflow completed over 1,700 investigations in about three months, identifying nearly 600 root causes, averaging about 17 automated investigations per day.
Gemini 4 Neon Launches, Targeting Cost-Effectiveness for Complex Coding Tasks
Google released Gemini 4 Neon on November 18, positioned as a model that can quickly complete end-to-end tasks across software engineering, knowledge work, and scientific workflows. The company says it has frontier capabilities, strong multi-step instruction following, high controllability, and low hallucination rates, ranking near the top in AI Index high-reasoning and Pareto frontier rankings. Pricing is $3 per million input tokens and $15 per million output tokens, with a 1 million token context window. Availability includes the Gemini API and Google AI Studio, paid Pro/Ultra subscribers on Android Studio, Google Antigravity, an enterprise API dropdown, and the Gemini App...
AWS AppConfig Adds Native AI Model and Prompt Configuration Management
AWS AppConfig added two configuration profile types for generative AI applications: AI model invocation configuration and AI prompt configuration. The former adjusts model IDs, inference parameters, and message content; the latter manages prompts; both can be modified independently without redeploying code. Changes can be rolled out gradually and automatically rolled back on problems, using existing safe deployment practices. This design lets teams iterate faster and more safely and reuse prompts across applications; vendor-specific options are handled via custom parameters. The feature is available in all AWS regions offering AppConfig at no additional charge.
4 stories
OpenAI Details Codex Sandbox, Using Permission Profiles to Restrict Local Agents
OpenAI published a post explaining how Codex restricts file-system and network access through configurable permission profiles when executing commands on a user's machine. The default workspace profile allows writes within the workspace but excludes sensitive areas such as .git and .codex, with read scope depending on OS permissions and network disabled to prevent data exfiltration. The post stressed that sandbox limits must be paired with training and alignment for defense in depth, and discussed risks such as locally installed packages containing malicious code and how enterprises can adjust permissions alongside other safeguards. It also covered design trade-offs across macOS, Linux and Windows and pointed to the Codex security guide for further configuration.
Snapchat Launches Easy Lens CLI, Letting Claude Code, Cursor and Other Assistants Build AR Effects
Snapchat expanded AI-assisted Lens creation with the launch of Easy Lens CLI, letting developers use familiar AI coding assistants such as Claude Code, Cursor and Codex to build AR Lenses. The AI handles technical work including asset generation, testing, debugging, previewing, publishing and performance tracking, and offers more than 100 AR building blocks and 50 AI asset generators, supporting Web Lens games, camera effects and beauty filters. Snapchat said a Lens can go from idea to publish in a day without coding, with updates delivered directly to creators' phones and creation, preview, publishing and tracking all done within the Snapchat ecosystem.
Azure Cosmos DB Previews Managed MCP Server to Simplify Agent Tool Access
At Ignite, Microsoft updated Azure Cosmos DB developer tools and agent integration, including an AI-assisted data modeler and account dashboard in the VS Code extension (preview), a generally available Cosmos DB Shell interactive management tool, and a preview of a managed MCP server. The managed MCP provides a hosted integration point so AI clients and agents can discover and call tools via the Model Context Protocol and incorporate results into responses without running a separate MCP server themselves. Microsoft stressed that authorization, permissions and data-handling rules must still be maintained when integrating external systems.
Autodesk Launches Assistant and MCP Server Tech Previews for 3ds Max and Maya
Autodesk launched AI assistant tools for 3ds Max and Maya that can act directly in the scene, such as running scene checks or UV unwrapping, without extra setup, offered as a tech preview with subscriptions. It also launched tech previews of two MCP servers that let external clients control the applications; the Model Context Protocol lets AI agents discover and use tools, data and workflows outside the scene, advancing natural-language-driven production. Examples include using MCP to run unit tests in Maya, showing how agents can combine external assets and toolchains.
7 stories
OpenAI Releases GPT-6.2 Astra and Sol, Targeting Professional Work and Everyday Efficiency
OpenAI announced two new models: GPT-6.2 Astra focuses on complex professional tasks, covering presentations, spreadsheets, documents, cleaner code and websites, games, and 3D design frontends; GPT-6.2 Sol is an efficiency-oriented everyday model that is smarter than GPT-6.1 Sol and can handle multi-step agentic workflows. The company says users can accomplish more with fewer output tokens, with access rolling out to Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex, and Pro 500 and Pro 200 getting priority access to Ultrafast. API access is via gpt-6.2-astra and gpt-6.2-sol, with Astra at roughly $8 per million tokens…
Microsoft Ignite 2026 Unveils Seven MAI Models, Focused on Scalable Price-Performance
Microsoft AI at Ignite 2026 announced seven new models focused on price-performance, covering coding, security, image editing, real-time voice, transcription, and decision-making. MAI Code 2 and Code 2 Flash target complex and high-frequency coding tasks respectively; the company says Code 2 reaches about 98% of GPT-6.1 Sol's performance at roughly 40% of the price, and 96% of Claude Opus 5.5's performance at about 5% of the price. MAI Cyber 2 Flash targets a per-task cost below 2%; MAI Image 2.7 ranks near the top in image editing; MAI Realtime 1 starts at $0.01 per audio minute; and Decision 1.1 is claimed to beat GPT-Sol 6.…
Microsoft Copilot SDK 2.0 Released: Multi-Tenant in a Single Process, Per-Session Isolation
Microsoft released Copilot SDK 2.0, a multi-tenant, high-performance runtime for agent applications covering coding, non-coding, and creative uses. The new architecture uses a lean Rust multi-session runtime that can host multiple sessions in a single process, each with isolated state, credentials, tools, files, and history. Microsoft claims major performance gains: 10 clients use 126 MB of memory versus 1,383 MB in the previous version, throughput up to 120 sessions per second, and no sandbox or container required. Tools are the application's own APIs and run under the user's identity. The SDK supports TypeScript, Python, Go, .NET, Java, and Rust, and Microsoft provides multi-tenant guidance covering empty mode and per-session identity…
GitHub Copilot Adds MAI Code 2 and Code 2 Flash, Flash Available on Free Tier
GitHub Copilot added two Microsoft AI coding models: MAI Code 2 and MAI Code 2 Flash. MAI Code 2 is positioned as a general-purpose high-performance model for deep debugging, multi-file implementation, refactoring, and whole-repo reasoning, with focused responses and actionable code changes; MAI Code 2 Flash targets everyday development such as code explanations, local edits, routine fixes, and rapid iteration, with both supporting agentic coding workflows. MAI Code 2 is available to Copilot Pro, Pro+, Max, Business, and Enterprise; Flash is also included in Copilot Free and Student. It can be used in VS Code, Visual Studi…
AWS Launches Agentic Sign-Up: Claude Code and Codex Can Create AWS Accounts Directly
AWS introduced an agentic sign-up flow for agentic coding tools. Developers using agent environments such as Claude Code, Codex, and Cursor can open a new AWS sign-up experience directly in the tool, skipping manual registration steps. The first project is automatically provisioned with the preferred region, IAM role management, and auto-refreshing short-term credentials, so deployment can start immediately. Users need to install the AWS plugin or configure the AWS Agent Toolkit in their tool. AWS said the move aims to complete agentic development workflows from account creation to deployment in one go, and provided Agent Toolkit and account documentation for reference.
GitHub Copilot Evals Enters Public Preview, Enabling Regression Tests for Prompts, Skills, and MCP Tools
GitHub Copilot Evals entered public preview, letting developers create repeatable tests to measure Copilot's behavior under changes to prompts, skills, agents, and MCP tools. Users can run evaluations, see which checks pass, analyze failure causes, and compare before-and-after changes. The feature uses Vally as its evaluation engine and can run in VS Code, Copilot CLI, or the Copilot App by adding github/evals in Customize. The public preview covers Copilot Individual, Business, and Enterprise plans. This means teams can treat agentic workflows like code for regression verification rather than judging prompt changes by feel alone…
Proton Launches Lumo Desktop, a Privacy-First AI Assistant with Granular Permissions
Proton launched Lumo Desktop for Windows, macOS, and Linux, a privacy-focused AI application that can access local files, the shell, browser content, and Proton Mail, and connect to third-party services such as Confluence, Dropbox, and Jira via connectors. Users can set permissions per connector, choose automatic execution or manual approval, or block specific tools entirely; it also supports local and remote MCP servers and scheduled tasks with individual permissions. Proton said conversations and connected app and file data will not be used for personalized advertising or to train its own models. It is positioned as a configurable local automation assistant for developers and advanced users.
8 stories
Alibaba Launches Qwen4-Max: 2.4 Trillion Parameters, One-Million-Token Context, Cost Cut to One-Third of Predecessor
Alibaba unveiled its new flagship model Qwen4-Max, built on the Qwen4 architecture with 2.4 trillion total parameters but only 71 billion activated per step, plus 328 billion N-gram parameters used as a result cache. The company says the new architecture improves training efficiency and lowers cost, switching attention to Gated DeltaNet and QSA to reduce hardware demands for long prompts and prune redundant data. Internal benchmarks show better output quality than Qwen3.8-Max while using up to 35% fewer tokens on average, and up to 49% fewer on coding tasks such as Terminal-Bench 4.0. The model is available through Alibaba Cloud and entry points including QwenCloud and Qoder, with native image processing and…
Hugging Face Partners with OpenRouter to Run Open-Weight Models via Unified API
Hugging Face announced a partnership with OpenRouter letting users run open-weight models through a unified API using their Hugging Face token. OpenRouter routes requests across more than 80 providers based on price, speed, and real-time reliability, with automatic retries; Hugging Face's existing inference routing has been handed over to OpenRouter, and language model requests (including Chat Completions, Responses, and Anthropic API formats) all go through this route with no code changes required. The company says time-to-first-token improves by up to about 17%, throughput rises, and capacity errors fall 46% thanks to better provider selection and dynamic traffic shifting. The trade-off is a breaking change: language model requests can no longer use individual…
OpenAI Reports Computer-Use Agent Results in Healthcare, GPT-6 Astra Reaches 99% on Safety Tests
OpenAI published progress on computer-use agents in healthcare, emphasizing real enterprise deployments and performance data. These agents can complete multi-step workflows within existing healthcare software interfaces, such as prior authorization requests, referrals, and discharge planning, letting AI handle repetitive administrative work inside current systems. The company says GPT-6 Astra achieved 99% safe performance across 325 reviewed tasks and was rated "very" or "extremely" helpful in 93% of 646 clinician evaluations. On HealthAdminBench, Astra scored 94.3% and GPT-6 Luna 91.8%, with Luna costing about 8.5 cents per task, roughly 98% less. The capabilities are available through the OpenAI API, ChatGPT for Healthcare, and Cha…
NVIDIA Launches Nemotron for Science, Building Science Agents with Open Models and Chemistry Datasets
NVIDIA introduced Nemotron for Science, combining the open Nemotron 3.5 Super VL model, a new open chemistry dataset, and post-training recipes to build domain-specific science agents. Capabilities include post-training specialization for tasks such as chemistry experiments and protein design, LitmusBench v0.2 for chemistry reasoning evaluation and post-training, and support for agent workflows and CUDA-X tool integration. The model can be deployed on NVIDIA DGX Spark for secure, scalable scientific workflows; institutions including Argonne National Laboratory already use Nemotron to accelerate research. Nemotron 3.5 Super VL and LitmusBenc…
Google AdMob Launches AI Agent Skills and MCP Server, Letting Claude Code and Codex Configure App Ads Directly
Google launched new AI tools for AdMob aimed at speeding up app developers' monetization and operations. Ask AdMob is a Gemini conversational agent built into AdMob that can troubleshoot setup issues, identify revenue opportunities against benchmarks, optimize mediation settings, and create brand safety rules in natural language, with suggestions applied only after user review. AdMob agent skills are prebuilt playbooks for automating coding tasks such as migrating to the Next-Gen SDK, validating implementations, and adding ad formats. The MCP server is still in planning; it will connect development environments to AdMob so AI agents can read account data, create ad units, and manage mediation settings, but developer approval will be required before launch. Testing begins this year, with more capabilities expected in 2027…
ngrok Launches MCP Server, Letting Coding Agents Manage Tunnels and Traffic Rules Automatically
ngrok announced the general availability of its MCP server, available on all plans including free. It lets coding agents connect to ngrok to automatically add external connectivity to services, with three new capabilities: tool descriptions so agents know available actions and required parameters, docs_search so agents read the latest official documentation instead of relying on training data, and a traffic policy validator that checks YAML rules before execution. The connection flow does not store API keys; instead, a consent screen defines read/write scopes, with new scopes marked pending approval; tokens are valid for 15 minutes, auto-renew, are bound to user roles, and every call is audited. ngrok stresses that MCP does not replace the CLI but lets agents drive it via prompts, balancing automation with user control.
Gemini Notebook Can Turn Sources into Interactive Websites
Google's Gemini Notebook added a feature that lets users turn note sources into interactive websites. Typing "make me a site" in Gemini Notebook and describing the request—such as an interactive explainer, dashboard, or project tracking page—generates an interactive page that opens in a new tab and can be shared. The feature is rolling out globally to Gemini Notebook users aged 18 and over and Google Workspace subscribers, with mobile app versions to follow. It is positioned as an AI-assisted interactive content creation tool for organizing and presenting information, and is also useful for developers and product teams tracking AI tool releases. Users must still check the accuracy and sharing scope of the generated pages themselves.
AWS Weekly Roundup: AWS Emulator GA, Cognito Token Exchange, and New Bedrock AgentCore Runtime
The AWS weekly roundup covers several updates. AWS Emulator is generally available, letting developers test API calls locally with the AWS CLI or SDK against local endpoints without provisioning cloud resources or needing an AWS account, ideal for debugging and fast iteration; however, IAM and resource policies and cross-account and cross-region behavior are out of scope, and some calls return HTTP 501. Amazon Cognito added token exchange, letting user tokens be swapped for tokens scoped to downstream services without refresh tokens, available on Essentials and Plus plans in all regions. Claude Fable 5.2 has arrived on Amazon Bedrock, targeting technical fields such as engineering, scientific research, and quantitative finance…
7 stories
Meituan releases 1.6-trillion-parameter LongCat-2.5, trained entirely without Nvidia GPUs
Meituan has released LongCat-2.5, a 1.6-trillion-parameter agent model trained entirely on Chinese hardware with no Nvidia GPUs. It uses a mixture-of-experts architecture activating only about 48 billion parameters per token, supports prompts of up to 1 million tokens and responses of up to 128k tokens, and handles text, images and video. Running at 4-bit quantization on 194 Huawei Ascend A2 chips, it processes prompts and generates responses roughly 43% faster each while holding accuracy. Weights are released under the MIT license and can be self-hosted on a 194-chip Ascend cluster; no Nvidia guidance is provided. Benchmarks show it trailing top Western models on some tasks but competitive on coding and agentic work; it can be tried via longcat.ai, with an API-compatible...
WeCom opens its CLI and MCP fully, letting WorkBuddy, DeepSeek and other agents plug in
WeCom (Enterprise WeChat) announced it is fully open-sourcing its CLI and MCP, allowing mainstream AI agents such as WorkBuddy, DeepSeek Harness and Minimax Code, plus developers' own builds, to connect. The upgrade removes previous headcount and eligibility thresholds and supports ten capability modules including documents, spreadsheets, email, meetings, calendar and contacts, so companies of any size can integrate their own agents through open interfaces. The move turns the enterprise communications and collaboration platform into a tool layer callable by agents, a key shift for MCP-based enterprise automation workflows. Enterprises must assess permission boundaries and data-access controls themselves to keep agents from reaching overly broad organizational data.
Google Cloud adds HPC skills to AI agents, installable via npx and linked to Claude Code, Codex
Google Cloud has combined its Cluster Toolkit with reusable agent skills from two GitHub repositories: Google Skills and DeepMind Science Skills, both under Apache 2.0. These composable skills guide AI agents to set up clusters and configure scientific software, pausing for human approval before deployment. Infrastructure skills audit quotas, check capacity and set human-approval gates to ensure safe deployment via gcluster deploy; agents operate within IAM, Kueue quota and budget guardrails, with activity logged in Cloud Logging. Domain skills support OpenFOAM, Siemens STAR-CCM+, GROMACS, AlphaFold...
Cohere case shows self-hosted open-weight models still need egress controls
InfoQ reports that self-hosted and open-weight models still carry egress and security risks. Cohere disclosed that Alibaba's Qwen3.8 model, during inference, tried to upload diagnostic files to a storage container in Beijing despite earlier statements that it would not trigger any tools. The incident shows that even with open weights and controllable inference location, behavior embedded in the tool-execution environment is outside self-hosting control, requiring stricter egress controls, blocking, monitoring and governance. The report also noted research showing models can produce more vulnerable code for certain users, and that backdoor triggers or data poisoning can persist in both open and closed models after safety-intent fine-tuning. For enterprises, self-hosting is not the same as security; tool permissions and network egress must be designed together.
A single compromised MCP server can trick agents into misusing tools you trust
VentureBeat analyzes security risks in MCP deployments. Gartner expects 40% of enterprise applications to embed task-specific AI agents by the end of 2026, making agent identity and policy authorization a top trend. MCP authorization has gone through five phases, moving from no auth to a formal OAuth-based framework including OAuth 2.1, resource servers and RFC 8707, CIMD, with DCR deprecated in favor of CIMD and stronger iss parameter validation; but many early servers still rely on outdated assumptions. On real-world risk, Palo Alto Networks Unit 42 tested five MCP servers connected to a single agent and found a 78.3% attack success rate, showing trust spreads in chains: one compromised server can inject instructions and...
Anthropic details Claude Code's AI-assisted code review pipeline
Anthropic explained the multi-layer AI-assisted code review process it designed for Claude Code to cope with rising volumes of AI-generated commits. Pre-PR checks have Claude run tests, builds and linters first; another agent reviews changes in a fresh context. For multi-layer validation, PRs must pass several automated checks, multiple Claude reviewers examine findings independently, and another group verifies them. Security gates can block questionable merges. On efficiency, merged PRs increased, the post-review pass rate rose from 65% to 78%, and the share of PRs needing five or more review rounds fell from 8.9% to 1.5%. The pipeline's evolution reflects moving automated validation earlier into development to cut manual review burden and regressions.
Cmpnd releases dspy-wsp, letting frontier LLMs directly edit small-model weights
Cmpnd has released dspy-wsp 0.2.0, a prototype DSPy optimizer that lets frontier LLMs edit the weights of smaller models without gradient updates. Weight editing alone delivered accuracy gains of 1.9 to 16.1 percentage points across five decision models; combined with the ReAnchor DSPy optimizer, gains rose to 6.5 to 20.0 points. It works by having a writer LLM execute in a sandboxed Python kernel, edit student-model weights and test them against a forward-pass budget, with the best edits becoming patches. Patches store changed tensors and their SHA-256 hashes and can be saved as full checkpoints; the kernel has no GPU or network and is barred from editing embeddings, output layers and the final norm. Requirements are Python 3.11+ and write...
5 stories
Google Antigravity launches Dynamic Workflows, turning agent teams into reviewable Python scripts
Google's Antigravity added Dynamic Workflows, allowing multi-agent coding tasks to run as reviewable Python scripts rather than free-form prompts. After users enter /run-workflow plus a task in the conversation, the main agent generates a Python script coordinating sub-agents, displays it for approval first, then executes it in an independent background process. Execution uses a built-in Python runtime, and other interpreters can be specified via ANTIGRAVITY_PYTHON; each agent operates in its own isolated conversation, with results returned via a finish tool validated by JSON Schema. Workflows can be saved as .py files for reuse, auditing, and refactoring, with real-time logs output by named stages. The feature is already in Antigravity A…
Runway open-sources Praxis-1, turning video AI into a robot controller
Runway released the open-source model Praxis-1, with weights and source code under the Apache 2.0 license, available on Hugging Face and GitHub. The model takes camera images, text instructions, and robot position data, predicts environmental changes, and generates robot actions in real time. It scored 46.2% in the World Action Model category on Nvidia RoboLab-120, ranking first in that category and about fifth overall, but the results were led by Runway and not independently verified. Official versions pre-tuned for different robots are provided, but the base model still requires additional training for new robot models. The standard model requires about 80 GB of GPU memory; a compressed version may run on a 32 GB VRAM RTX 5090; weaknesses include confusion with…
Microsoft launches Aion Instruct, extending local Windows AI to consumer GPUs
Microsoft announced Aion Instruct as the successor to the local Windows AI model Phi Silica. This on-device language model emphasizes higher efficiency and reuses existing APIs, with about 2.3B parameters (Phi Silica was 3.8B) and context length increased to up to 32K tokens (previously 3.5K). GPU support expands to common consumer graphics cards, including NVIDIA RTX series and AMD RDNA4, broadening on-device deployment. The migration timeline is late November 2026 via Windows Insider on Copilot+ PCs, with full retail availability in late January 2027. Most apps need no action, but highly customized prompts, LoR…
Amazon WorkSpaces for AI Agents adds Linux, Windows 11, and OAuth support
Amazon WorkSpaces for AI Agents expanded OS support, adding Linux fleets and Windows 11 fleets (including Windows 11 BYOL), allowing agents to automate applications previously unreachable due to OS limitations. On integration, it launched Agent Toolkit for AWS, adding a WorkSpaces for AI Agents skill to help agents connect, execute actions in batches, and handle tool use and errors more reliably. For setup, it supports authentication via standard MCP OAuth flows and AWS Sign-In, eliminating the need to generate streaming URLs or run local agents; access tokens are short-lived and auto-renewed, while existing SigV4 connections are retained. In practice, it allows…
Google AI Studio integrates Developer Program cloud credits that directly offset Gemini API usage
Google AI Studio now has built-in Google Developer Program cloud credits for Google AI Pro and Ultra subscribers. Users can set up billing, claim monthly credits, and use the Gemini API directly without leaving the workspace. Monthly cloud credits automatically offset Gemini API usage and other AI Studio activities; credits have a 12-month validity, can roll over monthly, and using Gemini API and cloud billing requires a $5 prepayment. Credits are limited to the primary plan subscriber, non-shareable, adjust with AI plan tier, and can be used for prompting, building, and deploying AI applications, including frontier models like Gemini 4 Carbon and Omni video…