Agent Patch Notes: a free weekly email with the AI agent releases that matter and every tier change on our S/A/B list. Issue #1 covered Windsurf becoming Devin Desktop plus 6 more updates. Goes out Fridays.
Amazon WorkSpaces for AI Agents now records a session replay: screenshots become a video with a plain-language transcript, saved in your own S3 bucket. Would watching the replay make you trust a desktop agent more, or would you still want approvals up front?
Browser Use 1.0.6 makes BU 4.0 the default model. If you never pass an llm, your agent now pays 15x the token price of bu-fast. Passing llm='bu-fast' keeps the old model and price. Is the newer model worth 15x for your browser tasks? https://t.co/09WOfg51lr
Devin can now build and host internal web apps for a team: a dashboard or review queue on a private URL behind Devin sign-in. If an update fails to start, the previous version keeps serving. Would you hand your internal tools to a coding agent?
Codex CLI 0.169.0 changed a default: codex exec now keeps project config and MCP servers off in untrusted directories unless you pass --trust.
Safer for cloned repos, one more flag for scripts.
Keep the new default, or add --trust everywhere?
https://t.co/LuE684ugjk
@kaiargus I can see the appeal: one reviewed script is far more manageable than dozens of calls. Being able to inspect the generated code before it runs is key.
Claude Code 2.1.323 adds an experimental Code Mode: instead of calling tools one by one, Claude writes JavaScript that calls them. It's opt-in, and organizations can turn it off.
Would you rather review the script or the individual tool calls?
Cursor Projects went multiplayer in beta: a shared board shows which teammate and which agent owns each task, plus its PR. Coordinators can message each other but treat that text as untrusted input.
Would you let your agent take orders from a coworker's?
https://t.co/xUwyByRa7v
n8n Labs released Nod, a free open-source agent that runs on your own computer with its own sandbox VM and browser.
Before it sends, pays, posts or deletes anything, a separate permission check confirms first.
Would you trust a desktop agent inside your Gmail and Slack?
OpenCode Browse lets agents drive your real browser, with your tabs, logins and extensions, then hand the tab back to you for 2FA or payments. Per-session site allowlists set the limits. Would you let an agent work in your logged-in browser? https://t.co/NmAg2pciLn
Kiro bots are now generally available in Kiro Crew: named, long-running agents with their own chat, notes memory and workspace, running on schedules you set in chat, like checking overnight build failures.
Would you let an agent triage your overnight build failures?
Claude Code 2.1.319 adds a setting that denies any tool call that would wait on a permission prompt, so the turn keeps going instead of stalling.
Would you rather your agent skip the action or stop and ask you?
https://t.co/U62EZAXJxg
CrewAI 1.15.36 adds experimental conversational Agent sessions, and crewai run prompts for a missing API key instead of failing. Conversational sessions or keep crews fire-and-forget? #AIAgents
Vercel Labs agent-desktop 1.2.0 ships MCP: agents click and type via the OS accessibility tree with JSON element refs, no screenshots. Accessibility tree or screenshot computer use?
https://t.co/BuoIUdOp1F
#AIAgents
CrowdStrike added an AI Kill Switch in Falcon Guardian: security teams can terminate a rogue internal agent where it runs, block relaunch, and cut its network access. Human trigger, full audit trail. Auto-kill or keep a person on the switch? #AIAgents
DeepSeek open-sourced dsh-superagent: watches a GitHub repo, does the work, can reply with a PR. Research prototype or production path?
https://t.co/Woy8KeNhsK #AIAgents
@martin_valchev_ Building Ranked Agents: honest rankings and head-to-head comparisons of AI agents and agent tools, updated daily with new releases. https://t.co/D1orn3yHy7
GitHub Agentic Workflows is now generally available, and Factory just shipped Droid 1.0 with a new agent harness. Two big steps for agents doing real dev work.
Full release log: https://t.co/JQl2wwP1Is
Congrats @GeminiApp! Gemini Deep Research landed A-tier on the Ranked Agents AI agent tier list 🔥 There's a free embeddable badge for your site or README if you want to show it off: https://t.co/D6vvr8o4PQ