Welcome to the 304th edition of Coding Jag brought to you by TestMu AI!👐
AI agents can now write your code, run your tests, and open your pull requests. This week LangChain argued they should stop doing it on your machines: give every agent its own computer, a tiny sandboxed machine that starts in under a second. The logic is hard to argue with. When agents share your infrastructure, one compromised package is all it takes.
Microsoft patched a record 622 flaws in one Tuesday. GPT-5.6 went from locked preview to available for everyone, and GitHub opened its Copilot desktop app to every plan, including the free one. Plus Selenium 4.46, a genuinely useful VS Code update, and Cucumber co-founder Matt Wynne on harness engineering.
📬 Come across something useful or interesting? Just reply and let's exchange ideas.
News
08 min
openai.com
🔓 Two weeks after locking GPT-5.6 behind a 20-partner preview, OpenAI has opened it to everyone. The family comes in three sizes: Sol, the flagship; Terra for everyday work; and Luna, the cheapest at $1 per million input tokens. All three are live across ChatGPT, Codex, and the API, and Sol now tops the Artificial Analysis coding leaderboard while using less than half the output tokens of Fable 5. If you build or test with AI, your options just got wider and cheaper.
07 mingithub.blog
🧰 GitHub just opened the Copilot desktop app to everyone, including the free plan, on macOS, Windows, and Linux. You can even bring your own API key and use it without any Copilot subscription at all. One catch for larger teams: on Business and Enterprise plans, an admin has to enable Copilot CLI first.
07 minthehackernews.com
🔐 Swati Khandelwal at The Hacker News reports Microsoft's biggest Patch Tuesday ever: 622 fixes in one go, more than triple the previous record. Two flaws are already being exploited, including a SharePoint bug that lets attackers take control over the network without logging in. If your company runs SharePoint, this is the update to install first.
07 minai.meta.com
🤖 Meta Superintelligence Labs releases Muse Spark 1.1, a major upgrade to its flagship model. It actively manages a context window of 1 million tokens, uses a computer the way a person would by writing scripts when automation is faster and clicking when it is simpler, and can split big jobs across parallel subagents. Developers can build with it through the new Meta Model API, now in public preview, and it is live in the Meta AI app today.
AI
09 minlangchain.com
🖥️ Amy Ru at LangChain makes a simple case: an AI agent running code on your infrastructure is one exploit away from causing real damage. The fix is LangSmith Sandboxes, where every agent gets its own tiny virtual computer, with its own kernel, that starts in under a second. Recent scares back the argument, like the npm worm that slipped malicious code into hundreds of packages. A practical read if your team is giving agents more autonomy.
10 minmartinfowler.com
📐 Unmesh Joshi, Distinguished Engineer at Thoughtworks, argues the secret to reliable AI output is not better prompts. It is giving the model a small, strict language to write in. LLMs already nail Mermaid diagrams and SQL because those formats leave little room to go wrong, and he shows the same idea applied to testing distributed systems, where a badly formed test simply refuses to compile. A mindset shift worth stealing for your own AI workflows.
09 minblog.n8n.io
🤔 Ophir Prusak at n8n tackles a question many teams are quietly asking: when do you hand a job to a coding agent, and when do you build a proper workflow? His answer boils down to five questions about who decides, who is involved, and what happens when things break. The rule of thumb: if the stakes are low and you are watching it run, use Claude Code. If it is recurring, unattended, and has real consequences, build a workflow.
Automation
07 minselenium.dev
⚒️ Diego Molina announces Selenium 4.46, and the headline fix is security: Selenium Manager now blocks a path-traversal attack when unpacking archives. Java users get a stricter, standards-compliant JSON parser, and BiDi support now works with Safari options. A small release on paper, but the security fix alone makes it worth updating.
10 mintestmuai.com
🕸️ Samyak Goyal at TestMu AI maps out how multi-agent systems are actually wired: five coordination patterns, from simple pipelines to supervisor and debate setups. Each pattern fails in its own way, and the guide shows how to catch those failures by running specialized AI evaluators in parallel against auto-generated test scenarios. With Gartner predicting over 40% of agentic AI projects will be canceled by 2027, testing the orchestration layer is where quality teams can make the difference.
Tools
06 mincursor.com
🗂️ Cursor 3.11 fixes one of the more annoying parts of working with AI: losing your main thread every time a side question comes up. New side chats open with a quick /side command, inherit context from the main conversation, and can feed their findings back in when you are done. Add search across all your past agent conversations, and long-running AI work finally feels organized.
07 mincode.visualstudio.com
💬 VS Code 1.128 lands two upgrades your whole team will actually use. Chat now accepts images and PDFs, so you can paste a screenshot of a bug and ask about it directly. And multi-chat sessions let you run conversations in parallel, compare approaches, and branch from an earlier point. A quiet release that makes everyday AI-assisted work noticeably smoother.
Video & Podcast
08 mintestguild.com
🎤 Joe Colantonio hosts Matt Wynne, co-founder of Cucumber, on what he calls harness engineering: building the structure around AI instead of just prompting it better. Wynne's core point is that shared understanding is still the real bottleneck, even when agents write the code. He also shares how teams use multiple LLMs to review AI-written pull requests, and what he learned modernizing decades-old mainframe systems with LLMs.
07 minyoutube.com
📺 OpenAI's official launch video shows ChatGPT Work in action: give it a goal, and it pulls context from your apps and files, breaks the job into steps, and delivers finished docs, sheets, and slides. It can stay with a complex project for hours. Worth ten minutes to see where this kind of hands-off AI is heading. Codex, which powers it, now counts more than 5 million weekly users.
Events
07 minministryoftesting.com
🎤 Ministry of Testing brings MoTaCon 2026 to the Brighton Dome on October 1: one day, one track, built for testers. The lineup already counts 24 contributors, including Lisa Crispin, David Burns, and Jason Huggins, with over 250 people registered. Entry comes standard with an MoT Professional Membership, and talks go up on the MoT site within a month.
07 mintestmuai.com
🎤 Clear your calendar for August 19 to 21. TestMu Conf returns as a free, virtual, three-day event with 80+ sessions on autonomous QA and AI testing, and 75,000+ testers and engineers expected from 120+ countries. Keynotes include Replit CTO Luis Hector Chavez and Entire CEO Thomas Dohmke, plus live challenges with prizes worth up to $15,000. Registration is free, so grab your spot.