AI news for Nov 19, 2026
21 stories
- DevelopersGoogleGoogle adds Tool Retrieval to the Gemini API for on-demand tool schemas
Google released Tool Retrieval in preview for the Gemini Interactions API, aimed at agents connected to dozens or hundreds of tools. Instead of sending every parameter schema with each request, developers add a tool_search tool and mark other tools with defer_loading: true. The model still sees each tool's name and description, then requests a schema by name and calls the function inside the same request.
- ResearchHacker NewsAnthropic says Claude solved two open black hole physics problems
Anthropic reports that Claude solved two open problems in theoretical physics concerning black holes and quantum gravity. One completed a calculation physicists outlined in 2003 about whether a phase transition remains abrupt in a regime where spacetime is not well defined. The other established an exact ceiling for a bootstrap method used to test candidate quantum gravity theories. Anthropic says the work came from agents running in the background for about six weeks, with results formally verified in Lean and main results checked by a physicist. The preprints are available. The company says neither result describes our universe directly or has practical applications.
- DevelopersOpenAI CodexOpenAI ships Codex 0.171.0 with GPT-6.2 Sol as default model
OpenAI published version 0.171.0 of Codex, its command-line coding agent. The release adds GPT-6.2 Sol and Astra to the model catalog, with Sol set as the default selection, and turns the terminal workspace on by default so document previews, image previews and subagent transcripts sit alongside the main conversation. It also adds configurable undo and redo shortcuts that restore attachments and large pastes, a codex exec --remote flag for starting or resuming noninteractive sessions on an existing app server, /rc and /remote commands for local remote control and device pairing codes, and /statusline for choosing MCP status line providers. Fixes cover drafts and queued follow-ups across session switches, permission profiles when changing working directories, and terminal scrolling and redraw glitches.
- ProductsAnthropic Claude CodeAnthropic ships Claude Code v2.1.329 with hook warnings and tasks view
Anthropic published version 2.1.329 of Claude Code, its command-line coding agent, adding visibility features alongside a long list of fixes. A new trust dialog now states where a folder's hooks are declared, and sessions in Claude Desktop and other non-interactive hosts warn when hooks from managed settings fail to load. The /tasks view now lists the session's scheduled tasks, /loop wakeups, the current goal and conversation forks, with x stopping a task after confirmation. /mcp gains a row showing JSON-RPC errors from remote HTTP servers that refuse a connection, and /status names managed marketplaces that receive no automatic updates. Code Mode can now run with OpenTelemetry events and metrics enabled.
- DevelopersGoogle ADKGoogle ADK 2.15 adds Maps tools, buffered sessions and pass^k evals
Google published version 2.15.0 of its Agent Development Kit for Python. The release adds a GoogleMapsToolset so agents can validate addresses, reverse-geocode coordinates and fetch Street View imagery, and gives BigQueryToolset's ask_data_insights a search_scope parameter for discovering tables across projects and datasets instead of naming them up front. Sessions gain an experimental BufferedSessionService that flushes events in background batches, and FsspecArtifactService now writes, deletes and versions artifacts across gs://, s3://, file:// and memory:// stores. Evaluation reports pass@k alongside pass^k for multi-run reliability and scores tool calls per argument. Sandboxes can be scoped per user or session, RAG memory search is faster, and ADK Web removes the interactive Visual Agent Builder.
- ProductsNVIDIA NewsroomNVIDIA Brings GeForce NOW to More Fire TV Sticks With Black Friday Discounts
GeForce NOW now runs on three Amazon Fire TV Stick models using Vega OS — the Fire TV Stick HD, Fire TV Stick 4K Select and the 2026 Fire TV Stick 4K — so members can stream supported games on a television with a connected controller and no console. NVIDIA paired the expansion with holiday pricing: 30% off the first six-month term of a Performance or Ultimate membership through Nov. 30, and 90% off the first month of a membership bought through Amazon on supported Fire TV devices through Dec. 31. Nine games joined the library this week, including GOG versions of five Bethesda titles for members who already own them.
- ProductsWindowsMicrosoft Edge launches Aion Instruct and opens Prompt API trial
Microsoft Edge announced that Aion Instruct, its 2.3-billion-parameter small language model, is generally available as an open-weights model on Hugging Face and is now Edge’s built-in on-device model, replacing Phi-4-mini. The company also opened an origin trial for the Prompt API, letting developers call on-device AI from live sites, and released a developer preview of the experimental Decisions API for structured yes/no, multiple-choice, and scoring answers in Edge Canary and Dev. Edge downloads the model only when needed and on capable devices; users can disable on-device AI in settings. The tools are aimed at web developers building hybrid on-device and cloud AI features.
- ProductsAWS Machine LearningAWS lets AI agents pay for Marketplace x402 products on the AWS bill
AWS has added AWS Billing as a payment option in Amazon Bedrock AgentCore payments, letting AI agents buy x402-enabled software, APIs, data and content listed on AWS Marketplace and charge the cost to the developer's AWS account. AWS also introduced an optional AWS Marketplace connector for AgentCore Gateway so agents can search listings and find seller endpoints on their own.
- ProductsAWS Machine LearningAmazon launches Selling Partner plugin for AWS Quick in beta
Amazon announced the Selling Partner plugin for AWS Quick at Amazon Accelerate on September 23, 2026, and it is now in beta for US stores. After one sign-in to a Seller Central account, the plugin connects AWS Quick to live listings, inventory and sales data, so sellers can ask questions in plain language instead of exporting reports. It installs like an app with no coding, and sellers grant None, View or Edit access per data category, which can be narrowed or revoked later. It ships with curated Amazon skills and pairs with Amazon Ads MCP. Separately, primary selling account holders worldwide can claim 12 months of AWS Quick Plus free, plus access for 10 delegated users, by signing up before December 31, 2026.
- ProductsAWS Machine LearningAmazon SageMaker AI adds endpoint monitoring for latency drift and idle GPUs
AWS has added monitoring to Amazon SageMaker AI that watches generative AI inference endpoints and flags when a change could improve performance or lower cost. Monitoring is on by default for endpoints running vLLM serving containers on GPU instances, with no setup and no added GPU cost, and customers can opt endpoints out. Working only from latency, GPU utilization and serving configuration — not request or response content — it reports latency drift, underused capacity, and configuration gaps such as a container missing newer optimizations or security patches. Findings arrive as one AWS Health notification per endpoint, which opens a prefilled session in SageMaker Studio where alternatives can be benchmarked on separate compute.
- ProductsAWS Machine LearningAmazon Bedrock AgentCore harness adopts Strands defaults and adds lifecycle hooks
AWS has updated the Amazon Bedrock AgentCore harness so its managed defaults now run on Strands, the company's open-source agent harness released earlier this year under the Apache 2.0 license. Agents that rely on the defaults get a recommended model, a task-oriented system prompt, prompt caching for Anthropic models on Amazon Bedrock, and summarization of older conversation turns once the context window passes a configurable threshold. New built-in tools cover task planning across invocations, web page retrieval, separate read, write and edit filesystem access, context offloading and delegation to subagents. The release also adds lifecycle hooks: Lambda functions can allow or deny an invocation or a tool call at four checkpoints, while Amazon SNS and EventBridge targets record activity. Existing harness users receive both updates with no new resource type to adopt.
- ProductsAWS Machine LearningAmazon Polly launches voice cloning for custom brand voices
Amazon Polly now offers voice cloning as a generally available, gated feature, AWS announced. Customers provide recordings from a consenting speaker, and Polly builds a custom voice that runs on its generative engine. The launch supports seven languages: English, Spanish, German, French, Italian, Portuguese, and Hindi, with each cloned voice speaking the language it was recorded in. Access is already enabled for select customers; others can contact their AWS account team or apply. Creating a voice is free, while use is billed pay-as-you-go. Cloned voices are encrypted, scoped to the creating AWS account, and generated audio carries a watermark. Customers can delete a voice through the API or console.
- ProductsAWS Machine LearningAWS adds cross-Region inference routing to Amazon SageMaker AI
AWS introduced cross-Region inference routing for Amazon SageMaker AI managed endpoints. The feature is built on a new resource called an inference router, which gives an application a single invocation path across endpoints and inference components in one or more AWS Regions and accounts. Customers register destinations for a model — such as a real-time endpoint or inference component — and invoke the router rather than a specific endpoint. The router keeps requests in its own Region while the local destination is healthy, shifts new requests elsewhere when it is not, and returns traffic gradually after recovery. Destination health is published to Amazon CloudWatch. Direct InvokeEndpoint calls still work, and the router does not create capacity, scale endpoints, or copy model artifacts.
- Open sourceCrewAICrewAI 1.15.44 adds versioned shareable artifacts
The update adds the ability to save files as versioned, shareable artifacts, and the documentation now shows flows how to save and read those artifacts. The release also includes bug fixes: inline agent and crew references now resolve relative to the flow file, nested defaults are isolated from shared flow declarations, image opening is limited to PNG, JPEG, GIF, and WEBP, and cache eviction order resets when cleared. Other fixes cover Anthropic file blocks, Claude 3+ vision detection, structured output from non-string guardrail results, and file path containment. The release is available now from the project’s GitHub repository.
- ProductsAWS Machine LearningPolyAI's Dialog-RSN-1 dialog model arrives on Amazon SageMaker JumpStart
PolyAI's Dialog-RSN-1, a model that reasons directly over live call audio rather than running a separate speech-to-text step, is now available on AWS Marketplace and in Amazon SageMaker JumpStart, AWS said. It is the first PolyAI model offered through AWS Marketplace. Customers can subscribe and deploy it as a SageMaker endpoint inside their own account — in their VPC, under their IAM roles, with CloudWatch logging — or launch it from the JumpStart catalog in one click or via the SageMaker Python SDK. The endpoint uses the OpenAI Realtime protocol, so an existing Realtime client connects through PolyAI's bridge with a base URL change. Audio must be 16-bit PCM mono at 16 or 24 kHz, sessions last up to 30 minutes, and the model is English-only today.
- DevelopersNVIDIA DeveloperNVIDIA adds packet reordering and low-latency polling to DAQIRI
NVIDIA released an update to DAQIRI, its Holoscan library for moving instrument and sensor data straight into GPU memory so measurements can be processed during acquisition instead of being written to disk first. The release adds hardware packet reordering offloaded to NVIDIA ConnectX-7 or newer NICs, an ibverbs engine that roughly doubles per-core throughput at packet sizes up to 1024 bytes compared with the previous DPDK backend in NVIDIA's benchmarks, and a direct polling mode with 2-4 µs send-to-receive latency in loopback tests. The post also describes a high-dynamic-range electron microscopy camera built at Lawrence Berkeley National Laboratory's Molecular Foundry that generates over 500 Gbps. Documentation, configuration examples and benchmarks are published.
- ProductsAWS Machine LearningAWS adds GPU diagnostics to Amazon SageMaker HyperPod
AWS introduced enhanced observability for Amazon SageMaker HyperPod, aimed at teams tracking down GPU problems in large distributed training jobs. The capability runs on clusters orchestrated by Amazon EKS or Slurm and is installed as an add-on or extension. It gathers default metrics every 30 seconds, with optional advanced metrics covering NVIDIA DCGM, AWS Neuron, host-level node exporter and Elastic Fabric Adapter network telemetry, sending them to CloudWatch, now the default store, or Amazon Managed Service for Prometheus. A diagnostic summary uses CloudWatch's DetectGroupOutliers API to flag GPUs that deviate from peers on utilization, memory, temperature, power and clock speed, while static rules catch XID errors, ECC memory faults, EFA throughput drops and thermal throttling, identifying the affected job, node and rank.
- ProductsDatabricksDatabricks opens Genie ZeroOps pipeline-fixing agent to public beta
Databricks has moved Genie ZeroOps, an autonomous agent for data pipeline troubleshooting, into public beta. The agent monitors jobs, pipelines and tables for hard failures as well as silent freshness and completeness problems, then pulls logs, source code, table metadata and lineage to find a root cause, tests a proposed fix in a sandbox, and sends alerts through Slack, Teams or PagerDuty. For Git-backed jobs it can open a pull request. Features added from private preview include shared inboxes, custom instructions that encode company workflows, and one-click PR creation. Workspace admins enable it from the Previews page. Databricks says the agent operates within Unity Catalog permissions and never changes production code or data without a person approving it.
- ResearchDatabricksDatabricks study finds Genie One accuracy rises with governed context
Databricks published a study testing its Genie One analytics agent on 125 questions about a fictional commercial auto insurer. The company compared four setups: tables alone, curated Genie Agents, a governed Unity Catalog layer without agents, and both together. Tables alone scored 64%; agents alone and the governed layer alone each scored 83%; combining curated agents with governed Metric Views, Domains, Subdomains and Pages reached 93–95%. Databricks says the two layers solve different problems: agents help with PDFs and documents, while governed definitions help with business terms. A retailer test raised accuracy from 54% to 82% when the governed layer was added to agents. Genie Ontology is generally available.
- ProductsDatabricksDatabricks makes bulk dashboard migration to AI/BI generally available
Databricks added bulk migration to Genie Code, letting teams convert a whole folder of Tableau and Power BI reports into AI/BI Dashboards with a single /importBI command. Users export reports to supported file formats, stage them in a Unity Catalog volume, then choose a target catalog and schema; the agent pairs workbooks with their data sources and builds dashboards and metric views. A status card marks each report as queued, running or completed, and runs can be stopped one at a time or as a batch. Databricks says the feature is generally available and requires partner-powered AI features to be enabled at the account and workspace levels. Each report gets its own conversation for review and refinement.
- PolicySalesforceSalesforce consortium selected for U.S. Super Intelligence export program
Salesforce announced that the U.S. Department of Commerce selected its consortium as one of 33 initial pre-set consortia in the American SI Exports Program. The designation covers a Salesforce package that U.S. and Foreign Commercial Service representatives and interagency partners will offer to foreign buyers. It also makes the package eligible for priority government-to-government engagement, export-control review and federal financing referrals. Salesforce said the package combines Hyperforce cloud infrastructure, Data 360 and Informatica for governed data, the Atlas Reasoning Engine with xGen and third-party models, Einstein Trust Layer and Salesforce Shield security controls, and Agentforce agents for service and operations. The program is intended to expand international markets for American technology.