Need to know
- Google's Gemini Agent gets its own Gmail address and routes tasks across Workspace, M365, and Slack as a universal work agent.
- Sophos cut threat investigation time 96% and resolved 52% of MDR cases end-to-end using OpenAI Daybreak AI agents.
- GitHub rewrote Copilot's runtime in Rust, dropping a key scenario from 5.25 seconds to 292 milliseconds.
- BeyondTrust patched a critical vulnerability in Remote Support and pushed a precautionary patch to Privileged Remote Access today.
- Asana cut browser-agent model costs 76x by optimizing on GPT-6.1 Sol using GPT-6 Astra in Codex to run experiments.
New Releases
Google announced Gemini Agent at its Gemini at Work 2026 event, a cloud-hosted universal work agent that receives its own Gmail address, calendar, and Drive storage and routes tasks across Google Workspace, Microsoft 365, and Slack.
- Model routing built in: the agent selects between Gemini and Anthropic Claude models per task, with OpenAI and open-weight models expected later, and includes cryptographic agent identity, role-based permissions, and audit trails.
- Governance layer included: an Agent Gateway controls traffic between agents and tools; real-time spend caps pause work when a project hits budget; private preview now, GA in North America expected by end of October.
GitHub replaced 800,000 lines of TypeScript and Node.js in the Copilot CLI, app, and SDK with a Rust runtime across 128 pull requests in 14.5 weeks without pausing releases.
- Startup time dropped from 5.25 seconds to 292 milliseconds when the Rust runtime runs embedded in-process, a reduction that compounds across every agent interaction at scale.
- AI-assisted rewrite: the migration itself was completed with AI tooling, making this a public proof point for using agents to do large-scale codebase modernization in production.
Microsoft Execution Containers, the sandboxed runtime that isolates AI agent permissions on Windows, Linux, and macOS, is now generally available.
- Four OS-isolated session types let administrators specify exactly which files, websites, and programs an agent can touch, with Windows enforcing those limits at the OS level rather than relying on model compliance.
- MIT-licensed and cross-platform: the code has been available for months but GA brings mature Windows 11 support and the policy controls enterprises need to deploy agents in regulated environments.
xAI rolled out personal email addresses for Grok Bot agents today, letting the agent sign up for services, contact businesses, and schedule meetings on a user's behalf.
- Addresses end in @mail.grokbot.com and are claimable by prompting Grok Bot or tagging @bot on X, giving each user's agent a persistent, actionable identity separate from the user's own inbox.
- Second major capability this week: the email launch follows Grok Bot's workflow-replication feature, in which the agent watches a user perform a task and then replicates it autonomously.
Funding
Nvidia-backed Upscale AI unveiled Token Fabric, a full-stack portfolio connecting heterogeneous XPUs inside racks and across cabinets — notable because Nvidia is funding infrastructure that explicitly helps customers mix competing accelerators, suggesting Nvidia sees ecosystem control as more valuable than GPU exclusivity.
Demis Hassabis announced $150M in investments supporting the White House and Department of Energy's Genesis Mission, a signal that frontier AI labs are moving from capability demonstrations to direct partnership with government science infrastructure.
Case Studies
Sophos deployed OpenAI Daybreak to cut threat investigation time 96% and resolve more than half of managed detection and response cases end-to-end without human analysts.
- Average AI response time is 89 seconds per case, replacing a workflow that previously required analyst triage, context gathering, and manual investigation steps that took multiples of that time.
- 52% of MDR cases are now fully resolved by AI agents, with the remainder escalated to human analysts — a ratio that directly addresses the security industry's analyst shortage problem at scale.
Asana used GPT-6 Astra running inside Codex to run optimization experiments on its StackAI browser agent workflow, then migrated the production workload to GPT-6.1 Sol, achieving a 76x cost reduction and 5x speed improvement.
- AI used to optimize AI: Astra in Codex ran the experiments that identified the cheaper model path, making this a compound case study where one model generation tunes the economics of another.
- Production browser agents at Asana's scale navigate websites, fill forms, and gather data for customers without code — the cost multiple means workflows that were previously economically marginal are now viable.
Trending on X
- Anthropic pulling ahead of OpenAI Developers on X are debating whether Anthropic has structurally overtaken OpenAI after releasing three competitive models in under three weeks, with sentiment running that the Claude lineup now covers every tier of OpenAI's product stack.
- Muse launched over safety objections The Verge's report that Zuckerberg accelerated Muse's launch after seeing Instinct gain traction — and despite internal safety concerns including an unauthorized password change — is generating significant reaction about whether competitive pressure is overriding safety process at major labs.
- Muse profitability math questioned Barron's analysis arguing that Meta Muse's token economics make profitability implausible at current pricing is circulating among investors and AI builders, with debate over whether Muse is a land-grab play or a structurally loss-making product.
- AI capex vs. capability growth diverging Francois Chollet's observation that AI capability gains fit an exponential from 2023 to 2026 while capex has been slightly super-exponential is prompting debate about whether returns on infrastructure investment are beginning to compress.
- Google's agent product fragmentation problem Ethan Mollick's argument that Google's Gemini 3 era was defined by fragmented products across many markets — and that this approach will fail in the orchestrator-agent era that Anthropic and OpenAI are defining — is being widely shared as context for evaluating the new Gemini Agent announcement.