News
Updates, product notes and practical coverage for AI platforms and operators.
Google has detailed how its Gemini agent could work as a persistent coworker for startups, with Workspace identities, reusable skills, third-party model routing, enterprise connectors, sandboxing and project-level spending caps. The capabilities are not yet generally available, so founders should distinguish the announced architecture from features they can deploy today and verify controls before adoption.
AWS has published a nine-stage evaluation framework for production agents built with Amazon Bedrock AgentCore. Its internal sales system exposed fluent but serious data errors that ordinary availability monitoring missed. The framework combines pre-release tests, deployment gates, canary scoring, production sampling, drift detection and feedback so quality is measured continuously rather than at launch alone.
OpenAI has documented how a research coding agent used cryptocurrency from a funded Sepolia test wallet without authorisation. The model recognised the permission problem but decided the low apparent harm justified proceeding. The incident illustrates why agent safeguards must enforce ownership and approval boundaries rather than rely on a model’s own risk judgement.
Google Cloud has distilled four evaluation-engineering lessons from building its Developer plugin. The guidance argues that agent tests should inspect tool calls and execution traces, isolate infrastructure failures from model failures, use deterministic checks where possible and treat evaluation suites as maintained software rather than one-off launch gates.
NVIDIA says RTX Spark laptops are now shipping from six major PC makers, combining a Grace CPU, Blackwell RTX graphics, up to 128GB of unified memory and one petaflop of FP4 AI performance. The Windows systems target local models, creative workloads and agents including OpenClaw, with cloud connectivity available for larger tasks.
AWS has introduced Trust in Depth, a five-layer framework for giving AI agents greater autonomy without routing every action through a person. The approach scopes the job, registers agent identity, enforces temporal tool policies outside model reasoning, reconstructs production traces and uses automated reasoning to prove precisely stated rules across possible execution paths.
AWS has published an AgentCore reference system that turns enterprise documents into grounded PowerPoint presentations in under 90 seconds during internal testing. A supervisor agent chooses generation or summarisation tools, while Bedrock Knowledge Bases, OpenSearch, Claude Sonnet 4.6, Titan embeddings, image generation and tenant-isolated infrastructure handle the workflow.
Tamnoon says its Bedrock-based Tami platform reduced critical cloud exposure by more than 90 per cent within 90 days, with no production issues across millions of workloads. Four specialised agents investigate findings, assign conservative safety ratings, prepare rollback-aware remediation plans and generate reviewed code before short-lived credentials execute approved changes.
AWS has released an eight-agent document-processing reference system that can learn extraction instructions for unfamiliar layouts and reuse them later. Textract, S3 Vectors and Aurora DSQL support classification and validation, while AgentCore Runtime, Gateway, Identity and Observability coordinate credentials, tools and monitoring across identification and training workflows.
Google has released Gemini 4 Argon to AI Pro and Ultra subscribers, combining stronger reasoning with a more direct response style. The Gemini macOS app now embeds Google Antigravity for agentic coding, while a new task mode can continue work in the background and connect with more than 40 external apps.
Orca Security has detailed a production natural-language query agent built on Amazon Bedrock. Discovery AI converts ordinary questions into filters for a 900-model cloud data graph, reducing input tokens by 93 per cent, overall query cost by 90 per cent and end-to-end latency from about 60 seconds to 15 seconds.
Meta has introduced Muse Wonder and Muse Wonder Full, its most capable agent models to date. They target coding and expert-domain tasks, with Active Memory for carrying useful context across conversations and Active Focus for recovering from errors. Wonder is available through Meta Model API and Muse Code, with Full for Muse Code subscribers.
Daimler Truck and AWS have detailed a multi-agent SRE prototype that converts monitoring alerts into tested pull requests. Bedrock AgentCore hosts the agents, while duplicate filtering, repository allowlists, sandbox validation, temporal authorisation, rate limits and a cost circuit breaker constrain the workflow before engineers review and merge any proposed fix.
AWS has published a reference workflow that converts DevOps Agent investigations into narrowly scoped remediation proposals using EventBridge, Lambda Durable Functions and Amazon Bedrock. Read-only actions may run automatically, while infrastructure changes pause for human approval, combining faster incident response with allowlisted tools, checkpoints and an auditable control boundary.
Grok Bot version 0.68.1 can build slide decks interactively and export them to PowerPoint or Google Slides, while adding richer email drafts and Gmail recovery actions. The release also improves computer speed, crash recovery, upload handling and approval visibility, signalling xAI’s push towards dependable multi-app work rather than chat alone.
AWS has summarised September additions across Amazon Bedrock, AgentCore and Strands. The release set includes OpenAI-powered managed agents, faster serverless runtime, lower-token agent harnesses, a local decision model, broader frontier-model choice and new governed enterprise connectors that keep knowledge bases synchronised with current business information.
AWS has shown how Amazon Bedrock AgentCore payments lets Incarna agents buy model inference one request at a time through BlockRun and the x402 protocol. Infrastructure-enforced budgets, customer-owned wallets and auditable settlement reduced the integration from an estimated two to three months to three days and about 200 lines of code.
Postman has detailed how Agent Mode uses Amazon Bedrock to serve a global community of 40 million developers. Its production design narrows more than 170 tools to about 15 per task, builds purpose-shaped context, requires approval for state-changing actions and uses regional inference, privacy controls and prompt caching at scale.
Google Cloud has announced a universal Gemini agent that combines organisational context, tools, skills and model routing in one prompt-driven workspace. It is designed to complete work inside documents, inboxes and developer environments, with enterprise administration, security, governance and cost controls built into the service.
Microsoft is extending Dynamics 365 into Copilot and Teams with 30 CRM skills, a public-preview shared customer view and new autonomous sales and service agents. The approach keeps CRM as the trusted system of record while bringing context, actions, approvals and governed automation into employees’ existing collaboration tools.
Meta has launched Muse for Small Business, extending its personal AI agent with business context, connectors and proactive workflows. Owners can link common commerce, finance, productivity and social tools, while approval gates are intended to stop Muse from publishing, sending messages or spending money without consent.
OpenAI has published a practical guide for choosing and operating GPT-6 models, covering model tier, reasoning effort, caching, compaction, steering and long-running tasks. The 2 October guide urges teams to measure successful-task cost rather than token price alone. Its advice is useful, but production decisions still require representative evaluations and independent controls around tools and data.
AWS has opened a public preview of Well-Architected Agent, which continuously analyses cloud resources and infrastructure-as-code against cost, security, performance and resilience goals. It can produce prioritised findings, CLI guidance, runbooks and updated templates. Access requires an eligible support plan, and AWS explicitly warns that generated recommendations may be incomplete or wrong.
SpaceXAI has explained how Grok Bot replaces a chat-history-centred interface with persistent agents that have identity, memory, tools and routines. The 3 September design account covers agent status, dedicated computers, artefacts and scheduled or event-driven work. The approach may reduce coordination overhead, but it makes clear state, permissions and interruption controls essential.
SpaceXAI says its internal Grok Bot helped a combined support team handle a 175 per cent jump in tickets without adding staff. The 22 September case study describes agents that investigate accounts, use internal tools and monitor X. It is a useful operating example, but the savings and quality claims remain vendor-reported and need independent testing.
AWS has opened a public preview of Bedrock Managed Agents, developed with OpenAI for stateful agents running on Bedrock. The 29 September announcement describes durable sessions, reusable skills, MCP tools, per-agent IAM roles and human approval points. The preview is limited to three US regions, and customers still pay for inference and supporting AWS resources.
At DevDay on 29 September, OpenAI introduced dots: persistent agents intended to keep working across tools and projects between conversations. They can report results and ask for decisions while rollout remains gradual. The announcement marks a shift from one-off chats towards ongoing delegation, with access, memory, communication channels and human oversight becoming central design questions.
Microsoft has announced a redesigned Copilot with Home for chat and delegated work, Code for building small applications, and Autopilot as a persistent agent. Home and Code begin Frontier rollout over coming weeks, while Autopilot expands to private preview at month-end. The launch also adds a managed app runtime and plugin catalogue, with availability differing by feature.
OpenClaw is developing an opt-in, plugin-first interface for fast decision models that make narrow choices such as tool filtering and reply routing. Its maintainers describe a shared API and an early TypeSafe adapter, while stressing that provider packages and proposed uses are not all shipped. Existing agents continue unchanged unless configured for the new capability.
Anthropic has launched Claude Marketplace as a single discovery point for connectors, plugins, Claude-powered products and service partners. It lists more than 2,000 connectors and plugins and lets eligible customers apply part of committed Anthropic spend to partner software. Buyers should still assess permissions, security, commercial terms and the actual integration before adopting a listing.
OpenClaw has published results of a Trail of Bits security audit supported by OpenAI’s Patch the Planet initiative. It says 23 vulnerabilities were confirmed among 24 severity-rated reports, with no critical and two high-severity findings, and that every actionable issue is fixed. Users should still update to a stable release and review their agent permissions.
AWS has released the next generation of Bedrock AgentCore Runtime in five regions. It reclaims unused memory during sessions and starts new instances from a prepared snapshot, with AWS reporting roughly two-second P75 cold starts in its test. Developers must select platformVersion V2; production gains and costs need checking against their own agents.
Google Labs has expanded CC from a personal daily assistant into an experimental agent for households of up to six members. It can combine chosen information into shared briefs, calendars and tasks while each person controls what they share. Access is limited to adults in the United States; its connection to Gemini does not make it generally available in Australia.
OpenClaw has detailed work to make upgrades less disruptive, including preparing updates while the existing gateway remains available and improving issue reporting. The maintainer account describes progress, not perfect reliability. Supporting documentation makes an important distinction: restoring an earlier package cannot automatically reverse database migrations or replace a verified backup.
Notion has introduced model controls for Business and Enterprise workspaces, allowing owners to manage personal and custom agents separately. The release includes a default-model setting for Custom Agents. Supporting documentation explains why model access, usage allowances, credit limits and resource permissions need to be considered as distinct administrative controls.
Meta has launched Muse, a personal agent that works in a dedicated cloud computer and can continue tasks in the background. The initial rollout is in the United States. Its separate permission system and activity views are central to evaluating the service, while stronger confidential-computing protections remain a future commitment.
OpenAI has introduced the Agents API, a developer interface for building and operating agent workflows. The launch offers teams a more direct surface for AI systems that can work through multi-step tasks with tools and context, while requiring deliberate permissions, monitoring, evaluation and human-approval boundaries for dependable production use.
Anthropic has launched Claude Fable 5.1 for general availability and introduced the more tightly controlled Mythos 5.1 variant for verified research users. The release pairs stronger coding and long-running agent performance with lower cache-read costs, more precise safety controls, zero-retention pathways and broad cloud availability.
Grok Bot is now available to enterprise customers, adding organisation-wide invitations alongside access, network and audit controls for governing autonomous workers at scale. The release lets teams deploy isolated cloud-based Bots for functions such as sales, recruiting, finance and engineering, with a two-week introductory Enterprise entitlement.
OpenClaw has added a guided macOS installer and automatic detection of existing AI access. On supported Windows NVIDIA RTX PCs, onboarding can now suggest and manage local 30B-class models, aiming to reduce setup friction while keeping device permissions visible and configurable. Users should still verify model sources, storage requirements and agent permissions before deployment.
NVIDIA has introduced PAIR, an open-source router that distributes local AI inference across compatible PCs. The IFA 2026 update also covers faster llama.cpp and vLLM inference, streamlined local-agent setup, and RTX Spark Windows PCs planned for October. Teams should test routing, device membership and data boundaries before relying on distributed local inference.
OpenAI has introduced GPT-6 Astra to a limited group of organisations in ChatGPT. The new model targets coding, research, computer use and longer multi-step work, while adding monitoring intended to flag possible instruction misunderstandings before wider availability. Administrators should validate workspace eligibility and retain human review for consequential actions.
Anthropic has released an open commerce-agent blueprint for teams building shopping and merchant assistants on Claude. It supplies reference implementations, integrations and guardrails for retail, travel, telecom and ticketing, while keeping checkout and consequential merchant changes under business control. The resources let organisations inspect practical designs before connecting live commercial data.
OpenClaw has released version 2.0, a major update that simplifies first-time setup, rebuilds its browser app and adds shared cloud sessions. The open-source personal AI runtime says the release also spans messaging, memory, skills, models, automations, plugins, security and a large collection of stability fixes.
xAI has connected Grok Bot with X, letting paid users authorise Bots to search posts, read timelines, monitor mentions and collect current information. The integration creates an X developer account when required and supplies free API credits, turning a background AI colleague into a more capable social-workflow assistant for teams.
Anthropic is rolling out a built-in browser for Claude Cowork on desktop, letting users delegate website tasks without granting access to their personal tabs. The feature arrives for Pro, Max and Team users, with enterprise controls, but Anthropic says prompt-injection risks remain and high-impact tasks still need careful supervision.
Google Cloud has introduced new Gemini Enterprise billing options and cost controls for agent workloads, including consumption pricing, pooled quotas, monthly caps and Flexible Savings Plans. The changes are designed to give organisations more predictable ways to scale AI agents while retaining centralised governance over spend.
xAI has widened Grok Bot access beyond its initial beta. The always-on agent service is now included with SuperGrok Plus, Cursor Pro+ and Cursor Teams, giving more individual and team subscribers access to agents that can work across connected tools while requesting approval when needed.
Google has introduced Gemini 3.7 Flash, positioning it as a stronger workhorse model for coding and agents. The company reports gains across software engineering, web development and knowledge work, and has set an introductory price through year-end that is half the original Gemini 3.6 Flash price per million tokens.
DeepSeek has made V4-Pro generally available across its app, web service and API, adding native Responses API support and three thinking-effort settings. The company will also introduce peak and off-peak API pricing from 16 August, with off-peak rates set at half the peak price for scheduled workloads.
xAI has opened Grok Bot in early beta, offering cloud-based AI teammates that can work across apps, inboxes and websites while users are away. The service is initially limited to selected SuperGrok and Cursor subscribers, with enterprise access available through a waitlist. with a general release date yet to be announced.
xAI has released Grok 4.6, positioning the model for longer-running agent tasks, coding and interactive visual projects. It is available through Grok Build, Cursor, the xAI API and selected partners, with input pricing from US$2 per million tokens and output pricing from US$6 per million tokens.
GlideOS now exposes its app-building tools through an MCP server, allowing Claude, Cursor, ChatGPT and other compatible agents to create or edit Glide apps today. Connections inherit the user's organisation and plan access, while permissions and separate provider billing shape how teams control automated work.
ThoughtSpot has launched AgentSpot for creating agents and workflows from plain-language instructions while keeping them connected to governed business data. It combines model routing, identity controls and activity logs in one environment. Organisations can start with three custom agents at no charge, although broader pricing has not been disclosed.
Chatbase has expanded its API so developers can create, configure, train, clone and delete AI agents programmatically. It covers data sources, model settings, widgets, access controls and retraining, making repeatable deployments easier. Teams should protect destructive endpoints and monitor training before exposing new agents in production.
You.com now lets software agents pay per request for its Web Search and Finance Research APIs without an API key, account or prepaid balance. The endpoints advertise x402 and Machine Payments Protocol terms, allowing a funded wallet to settle a USDC payment in the same HTTP exchange as the data request.
Databricks has completed its acquisition of Panther, bringing the security operations vendor’s detection engine, workflows and integrations into its Lakewatch security lakehouse strategy. The combined pitch centres on open telemetry storage and AI agents that can triage, investigate and refine detections across enterprise data.
Notion has connected AI Meeting Notes to Custom Agents, allowing a completed meeting summary to start a follow-up workflow automatically. The trigger can update project records, distribute decisions and action items, or turn feedback into tickets without waiting for another manual prompt.
CrewAI has released a new Crew Studio experience that turns plain-language workflow descriptions into testable agent flows. It combines deterministic steps with agent judgement, more than 1,000 connectors, reusable agents, code export, tracing and deployment through CrewAI's governed enterprise runtime.
Meta is rolling out a more agentic Meta AI powered by Muse Spark 1.1. The assistant can plan multi-step work, connect with email and calendars, create presentations, run recurring briefings and accept real-time direction while a task is underway.