---
title: Your agent pulled the night shift. Who clocks it out?
---

Plus: Developers are downgrading their AI on purpose ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­    ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­  

| View in browser Weekly Update  \|  Issue 545 Your agent pulled the night shift. Who clocks it out? Your AI agent can spend the sleeping hours rewriting GPU code, but after you wake up and have had your morning chaider, you still have to make the call: Is your AI-written code ready to ship?   So much of the conversation about coding agents glosses over the work they produce in favor of conversations about efficiency, speed, even luxury. But the uncomfortable question remains.   At Together AI, agents run experiments to improve the GPU kernels behind its inference service. The company says its compiler can squeeze weeks of tuning into an overnight run. But a human reviews the changes.   Raindrop is working to refine the answer to a question that comes earlier in the process: How do you know if a change made by your agent is safe to ship?   Then there's Google's donation of gVisor to the Cloud Native Computing Foundation. Anthropic and OpenAI already use gVisor to isolate code execution. Google hopes neutral governance will encourage broader participation, including from companies that need secure sandboxes … but can't afford to build that expertise themselves. For developers, these changes create a new type of approval-based work – alongside whatever time the agent saves.    As you catch up this Saturday, consider where your team would draw the line: What can an agent try, and what still needs your own signature as you sip your morning chaider?   — Nick Lucchesi, editor-in-chief, The New Stack 160 admin hours a year: how AI-guided diagnostics finds root causes faster A file transfer fails at 2 a.m. Somewhere in a pile of logs, configs, and partner connections is the reason, and finding it can take hours. Across a year, MFT admins can lose up to 160 of them. Join us live on November 19 and watch AI work through a real failure scenario. You’ll learn how to: Pinpoint the likely root cause with AI-guided diagnostics Catch problems early with anomaly detection Keep documentation current without writing it by hand For a deeper look, pre-register for our upcoming eBook, The Next Era of MFT, written by Carly Page and sponsored by Fortra. Save your spot TNS essential reads Developers are downgrading their AI on purpose — and it's working Anthropic added game-cloning limits to Claude Code, but developers say Opus 5.5 is now refusing reverse-engineering and software work it once handled. A developer opened Postman. Then two AI coding agents changed without their permission. A developer found Postman’s desktop app installed plugins into Claude Code and Codex without permission, reviving questions about coding agent control. OpenAI built itself a new Git backend, too Unlike Anthropic, OpenAI kept Git. Its new backend, Pando, cut push latency nearly 5x and serves close to a petabyte of Git data a day. From feral to fiduciary: making AI agents serve your interests Vijil argues AI agents should earn autonomy like credit: tested up front, capped by policy, and monitored constantly. Without that, they go feral. Nebius cut rollout collection time by 66.9% — here’s what changed In agentic RL, expensive GPUs often sit idle waiting on slow rollouts and cheap containers. Nebius’ Di Jin explains how to fix the bottleneck nobody owns. Kubernetes just formed a workload conformance group. Here’s why it matters for AI teams. Kubernetes is drafting a runtime standard for workloads as AI adds more variety. Meanwhile, agent security horror stories show what’s at stake. Why MCP servers should tell agents what to do next Most MCP servers still act like thin APIs: an agent asks for data, and the server hands it back. In this episode, AWS’s James Ward argues they should do more, guiding the agent’s next move with every response. He also shares AWS’s low-tech fix for MCP’s missing trace IDs: a session ID parameter that stitches individual tool calls into one journey, so teams can see where agents go wrong and measure whether their fixes actually help. Catch the episode Upcoming events & webinars Nov 10 Salt Lake City, UT Nothing but Net\[work\] \| KubeCon Reception with The New Stack and AWS ​As the cloud native community comes together for KubeCon, we are celebrating at the biggest party of the week in Salt Lake City... a Utah Jazz basketball game! Join The New Stack and our co-hosts, the AWS Open Source team, in the skyboxes for a night of elevated bites, beverages, banter, and, of course, basketball. Request to Join → Nov 11 Salt Lake City, UT Night at the Museum of Illusions with The New Stack and Dynatrace Bend your brain and expand your network by joining us at the Museum of Illusions to celebrate Kubernetes and the cloud native community. The brightest minds at the conference will mingle with VIPs and The New Stack's Editorial team, so don't miss this chance to connect. RSVP now → Nov 30 Las Vegas, NV re:Invent Reception with The New Stack & Hydrolix Join Hydrolix and The New Stack at AWS re:Invent for an evening of great conversations, meaningful connections, and signature cocktails at Delmonico Steakhouse, located inside The Venetian Resort in Las Vegas. Enjoy specially branded cocktails, catch up with familiar faces, and make new connections at one of Las Vegas’ premier steakhouses. Put your name on the list → Dec 2 Virtual Teaching AI agents to search like experts AI agents are only as good as what they retrieve. On the BrowseComp-Plus benchmark, GPT-4.1 answered 93% of questions correctly when handed the right documents, but just 15% when it had to find them with keyword search. Join us live on December 2 to see what agents need from search, and how to give it to them. Register now → From our sister site, Towards Data Science ShipAI: TDS's Video Showcase for Real AI Work Reading about AI is one thing — watching someone actually build with it is another. ShipAI is TDS's new archive of real AI projects: not a feed to scroll, but a record to keep, curated one entry at a time. See real, working AI builds instead of just written case studies Every entry is reviewed by an editor before it joins the index. Featured builds get pulled out each month as must-watch picks. Browse all projects TNS quote of the week “Most of what we build starts with our forward deployed engineers (FDEs), who work on-site with customers.” — Lukas Burger, who works across go-to-market and engineering at Cognition   Get the context → Connect with The New Stack The New Stack is a media platform for the people who build and manage software the world relies on. Sponsor this newsletter The New Stack 1111 6th Ave Ste 550, PMB 50938, San Diego, CA 92101-5211   Not loving everything we send? You can update your preferences or unsubscribe from all email communications. |
| --- |