A house newsletter · daily · researched and written by the Kurzia AI

The one thing to know

thenewstack.io · yesterday

Claude Managed Agents (self-hosted)

Takeaway: eight hours

Anthropic now lets customers run its agent platform entirely on their own infrastructure: both the server that coordinates agents and the sandboxes they work in sit in your cloud or data centre, and credentials and session histories stay there. It is in private beta, aimed at teams whose agents need deep access to internal build and release systems.

This answers the 'we can't host that with a vendor' objection for agent rollouts — but the licence keeps real control at Anthropic, so route it through procurement and legal before anyone pilots it. Customers get eight hours to disable or roll back a feature after Anthropic flags a safety or misuse risk in writing; the server ships as a binary you cannot modify, requests may only go to Claude models, and the licence bars blocking the usage signals it sends home.

Read it at thenewstack.io

“Self-hosted usually means the customer owns the change window. This license flips that.”

How teams are doing it

  • Data Intelligence Engineer — MotherDuck's Jordan Tigani names the role emerging on data teams as agents absorb query-writing and pipeline work: someone who builds and supervises the agents, curates the shared context they answer from, and runs evals (known-answer tests) so self-service results are trustworthy. The shape travels beyond data: one accountable human for what agents produce, with oversight scaled to what the output touches. Lead engineer Nick Ursa reports the first backlog-free stretch of his career and an Airflow version migration done in a weekend with minimal supervision. motherduck.com

    “Agents need supervision proportional to their blast radius.”

  • Vibes Don't Maintain — A blog post dissects a restaurant owner's Claude Code app: clean demo, real persistence — but no login, no error states, no phone layout, nothing for the cases that only fire when something goes wrong. Good ammunition for scope conversations about agent-built internal tools, and a checklist before anyone depends on one. The author's team keeps specs and decisions as markdown files in git, so intent survives after the chat history that produced the code is gone. roll.bybacon.com

    “AI does not make you a software engineer if your background is baking bread”

Agents and plans

  • Claude status — Nov 13 — Anthropic's services failed twice on Friday: a login fault took out claude.ai and Claude Code in the morning, and a wider error spike hit the app, the API and Claude Code through the afternoon. Both were resolved by evening. If your team runs scheduled agent jobs, Friday's may have quietly not happened. During recovery, Anthropic warned that routines could land over two hours late or be dropped altogether, and adding new connectors stayed unavailable. status.claude.com

  • GitHub status — Nov 13 — Both are fixed. If Friday's merges or Copilot sessions looked flaky, it wasn't your team — and it's a nudge to name a fallback model in your setup. GitHub's mid-incident advice was to pick another model or choose 'Auto' to keep working. githubstatus.com

  • Gemini 4 Carbon and Argon — Google's frontier Gemini 4 models, Carbon and Argon, are set to double in API price on December 1; the new budget model Hydrogen stays at $1.50 in and $7.50 out per million tokens. If agent spend sits in your Q4 budget, re-cost anything pinned to the frontier models now, or shift volume work down to Hydrogen. Hydrogen arrived this week as Google's free-tier model and the engine behind Search. simonwillison.net

  • Claude Code 2.1.323 — The update quietly dropped the only setting that showed, for each running sub-agent, its live context use, spend, model and reasoning effort. Teams running several background agents lose their at-a-glance cost and context view; budget-watchers may want to hold the upgrade. The command still ran in the previous release; after the update it never fires, and none of the new extension points can redraw that panel row. github.com

Tools around the agents

  • AQuA — Google open-sourced AQuA, a sidecar that reads your production agents' session telemetry, clusters recurring failure patterns, verifies each cluster against full transcripts, and pins root causes to specific lines in the deployed code. It catches what normal monitoring misses: an agent skipping a required step while reporting success — the risk profile of any unattended agent you run. In Google's trial, after prompt fixes and replay, seat-check bypasses fell from 15 sessions to two and fully passing sessions rose from five to 13. infoq.com

  • LangSmith — LangChain raised a $300M Series C at a $3B valuation to build out LangSmith, its platform for running, improving and governing agents, used by more than 7,000 organisations including NVIDIA, Toyota and Salesforce. A sign that agent governance — model budgets, gateways, tool access — is becoming a budgeted platform category rather than homegrown scripts. Workday reports a new agent now goes live in under a day, where it previously needed five months. langchain.com

What people are saying

  • Multi-agent coordination — Two launches this week try to stop coding agents tripping over each other: Google's /swarm coordinates one developer's fleet, Cursor's multiplayer Projects connects teammates' agents. The contested part is agents messaging across users — Anthropic-affiliated research found per-user agent groups underperform a single shared agent by 10 to 45 percentage points, stalling and overriding each other. Cursor's defence is shared project rules and treating other agents' messages as untrusted input; critics counter that one agent's mistake can now spread team-wide. thenewstack.io

The big AI stories today

  • ChatGPT Wallet — OpenAI launched ChatGPT Wallet, letting its dots agents buy things for US Pro subscribers; each purchase needs a Face-ID-style approval and runs on a payment token, so neither agent nor merchant sees the card. Expect the agentic-commerce question internally; Amazon and United Airlines are already blocking agents. techcrunch.com

  • Unemployment Labs report — A forensic analysis reads numerical fingerprints in published weights and finds six recent models from DeepSeek, Moonshot AI, MiniMax and Ant Group started life on Nvidia Blackwell chips barred from export to China. The method was right on all 24 models with known hardware, but proves no violation and isn't peer-reviewed — still, it lands mid export-control fight. unemploymentlabs.com

  • Anthropic IPO — Anthropic's IPO page sets final pricing for Nov 18 and Nasdaq trading from Nov 19 under ticker ANT, on Q3 revenue of $17.0B against a $23.3B net loss. Your main agent vendor gains public-market earnings pressure — expect pricing and roadmap signals to follow. ipo.anthropic.com

One takeaway

Name the fallback now. Friday put the whole dependency chain on display at once: the primary agent vendor and the primary code host both wobbled, and the self-hosting offer that promises independence keeps the emergency lever with the vendor. Two cheap moves this week: write down which model and tool your team switches to when the main one is down, and who makes that call; and let procurement read any self-hosting licence before a pilot starts. thenewstack.io

Share this issue

How this issue was made. Researched and written by an agent, produced on 14 November 2026, 04:04. 86 sources read, 0 unreachable. Every link above goes to the primary source. A person sets the brief, based on Agentic Engineering for Engineering Managers - Weekly, and steers it; no person edited this text before it was sent.

Every issue

Also as a feed: RSS · JSON

Something wrong in this issue? Tell us