Hermes Agent v0.16.0 "The Surface Release" Ships a Native Desktop App
NousResearch released Hermes Agent v0.16.0, dubbed "The Surface Release," the project's biggest interface overhaul to date. The headline feature is a genuinely native Electron-based desktop app for macOS, Linux, and Windows — not a terminal wrapper — with one-click installation, built-in auto-updates, drag-and-drop file support, clipboard image pasting, and a Cmd+K command palette. The app supports concurrent multi-profile sessions with cross-profile session linking, and can connect to a remote Hermes instance over an authenticated WebSocket (OAuth or username/password), letting a lightweight client talk to a more powerful server elsewhere.
The browser-based web dashboard has also grown into a full admin panel: configuring messaging channel integrations (Telegram, Discord, Slack), managing MCP server catalogs with simple enable/disable toggles, and handling credentials, webhooks, and memory settings can now all be done without SSHing into a server to hand-edit config files. A new "Quick Setup via Nous Portal" flow gets new users from install to first message in seconds, with a more detailed "Full Setup" path for advanced configuration.
Other notable changes include a fuzzy-searchable model picker available across desktop, web, TUI, and CLI (with new models DeepSeek-v4-flash, MiniMax-M3, and Qwen3.7-plus added to the catalog), a new /undo [N] command for reverting recent conversation turns, and the addition of NVIDIA/skills as a default trusted source in the Skills Hub alongside OpenAI, Anthropic, and Hugging Face. The release also patches CVE-2026-48710 (a Starlette BadHost vulnerability) and adds SSRF hardening. In total, the release closed 399 issues and merged 542 pull requests from 170 contributors.
Read more — NousResearch/hermes-agent on GitHub
OpenAI Codex Adds Branch/Worktree Support and a Browser Developer Mode
OpenAI shipped a cluster of Codex updates across its CLI, app, and mobile clients in early June. Codex CLI 0.139.0 adds a standalone "code mode" web search that returns plaintext results, preserves oneOf/allOf structures in tool schemas, and expands codex doctor to report editor and pager environment details for easier debugging.
On the app side, version 26.609 introduces the ability to choose a branch, create a worktree, and run an environment setup script when starting a new Codex thread — a meaningful workflow improvement for anyone juggling multiple in-flight changes against the same repository. The same release adds a "developer mode" for browser use with Chrome DevTools Protocol access, expands Computer Use to more enterprise regions, and adds per-app access controls for Computer Use on Windows. ChatGPT for iOS picked up matching branch/worktree selection, a new Codex profile screen with usage stats and token activity charts, /goal support for managing goals from mobile, and inline review comments on changed files.
Earlier releases in the window (26.602 and CLI 0.137.0) added activity insights and shareable profile cards, improved Computer Use startup reliability, F13–F24 keybinding support with paste in menus, enterprise monthly credit limit displays, and remote-control client pairing via app-server RPCs. Together, the updates push Codex further toward parity with terminal-and-IDE-centric agent workflows, particularly around managing multiple concurrent branches of work.
Read more — OpenAI Codex Changelog
GitHub Copilot Code Review Gets Org-Wide Runner and Content Exclusion Controls
GitHub shipped three administrative improvements to Copilot code review aimed at larger organizations. Admins can now set a default runner for Copilot code review that applies automatically across all repositories, with the option to lock that setting so individual repos can't override it — useful for teams standardizing on self-hosted or larger runners without repo-by-repo configuration.
Copilot code review now also respects content exclusion rules at the repository, organization, and enterprise level, so specified files or directories can be kept out of Copilot's review entirely — important for repos containing secrets, generated code, or vendored dependencies that shouldn't be sent to an AI reviewer. Finally, the 4,000-character cap on custom instruction files (.github/copilot-instructions.md and *.instructions.md) has been removed, allowing much more detailed project-specific review guidance.
Read more — GitHub Changelog
Cursor's Bugbot Drops to ~90-Second Reviews and Adds a Pre-Push /review Command
Cursor shipped a significant performance upgrade to Bugbot, its automated code review agent. Reviews now complete in roughly 90 seconds on average — down from a ~5-minute baseline, with 90% of runs finishing within three minutes — while also catching more bugs per run (0.62, up from 0.56) at about 22% lower cost per run. The gains come from Composer 2.5, the model now powering Bugbot, which Cursor says was trained on 25x more synthetic tasks than its predecessor. Composer 2.5 remains exclusive to Cursor's own products; on independent benchmarks like SWE-Bench Multilingual it trails Claude Opus 4.7 slightly, though Cursor's internal benchmarks show it ahead of competitors.
Alongside the speed-up, Cursor introduced a /review command that runs Bugbot and a security analysis locally before you open a pull request, available in Cursor 3.7+ and on cursor.com/agents (CLI support is coming). The command includes duplicate-diff detection: if you run /review locally and later open a PR with the same diff, Bugbot recognizes it, skips re-review, and posts a comment noting the prior analysis — avoiding duplicate charges for the same review. At Cursor's stated pricing of $1.00–$1.50 per run and 0.62 bugs caught on average, the cost per bug found works out to roughly $1.60–$2.40, which is trivial next to the cost of a production incident.
Read more — Digital Applied
