Google announced Gemini Agent (Oct. 8), a unified enterprise agent intended to run long-lived tasks from objectives, use Workspace/Microsoft 365/Slack/CLI and third-party surfaces, maintain persistent context, and spin up temporary subagents. Google says execution runs in cloud sandboxes with an Agent Gateway policy layer; this is a broad platform vision and availability/real-world performance remain to be established. Sources: https://cloud.google.com/blog/products/ai-machine-learning/welcome-to-gemini-at-work-2026/ ; https://news.ycombinator.com/item?id=50017609
GitHub made local sandboxing generally available for Copilot CLI, the Copilot app, and VS Code Agent Host sessions. The MXC-backed boundary controls filesystem, network, credentials, and local MCP/tool access across Windows, macOS, and Linux, with enterprise-enforced policies and no added Copilot charge—an important shift from optional agent safety preview toward deployable governance. Source: https://github.blog/changelog/2026-10-07-local-sandboxing-for-github-copilot-now-generally-available
Copilot for JetBrains added enterprise-managed default agent models and a switch to prevent automatic MCP server startup for Copilot and Claude; it also lets users jump from diagnostics into an agentic repair action. These are notable enterprise controls over model choice and tool activation, alongside reliability fixes for agent sessions and worktrees. Source: https://github.blog/changelog/2026-10-10-new-controls-and-chat-improvements-in-copilot-for-jetbrains
GitHub’s Oct. 12 limited public preview adds durable authenticated pusher identity independently of commit author/committer text, plus a ruleset to require committer metadata to match the authenticated pusher. This closes a spoofable identity gap that becomes more consequential as agents create and push commits; currently limited to GitHub Enterprise preview. Source: https://github.blog/changelog/2026-10-12-push-attribution-public-preview
Cloudflare expanded its newly introduced Clef decision-model family with Clef-omni (Oct. 9), which accepts audio and video as well as text and images, enabling one bounded structured decision model instead of cascaded transcription/vision pipelines. Cloudflare also made Clef faster and Clef-flash cheaper; open weights and Workers AI are available. Performance comparisons are vendor-reported, and the project is only days old. Sources: https://blog.cloudflare.com/clef-faster-cheaper-multimodal/ ; https://huggingface.co/Cloudflare/clef-omni ; https://news.ycombinator.com/item?id=50025236
JetBrains researchers used 4,782 real IDE agent sessions to argue coding-agent benchmarks should match the task flow of their intended users, including mixed code questions, planning, review, refactoring, and execution. In a 700-task pilot, splitting work into sequential interactions roughly doubled cost without stable resolution-rate gains; this is an author-reported study, but a useful production-grounded warning that interaction protocol can skew benchmark conclusions. Source: https://arxiv.org/abs/2610.09633