NuGet packages can ship agent skills alongside the library
On the .NET blog, Richard Lander shows that a NuGet package can include SKILL.md files, the open format for agent skills, with instructions for the agent to use that library the right way. The example is the markout package, which ships five skills.
Because dotnet restore already puts these files on disk, the documentation arrives in the same version of the package the project uses. The post shows how to list the skills with dotnet-inspect package markout -S Skills and how to copy them into .agents/skills with dotnet-package-skills install.
The text itself points to the security concern: a skill is an instruction for the agent, so the rule of only using packages from maintainers you trust applies. The author promises a second post on how agents consume these skills.
For .NET developers: library maintainers get a way to deliver documentation to agents in the same package version; consumers should review the skills the way they review any dependency.
Source: .NET Blog (Microsoft)
VS Code 1.145 brings remote agent sessions closer to the local experience
VS Code version 1.145 shipped on November 4 with a focus on agents. In preview, you can browse through the remote agent host connection (workbench.browser.enableRemoteProxy) and mix different harnesses in the same chat session (chat.agentHost.multipleChatProviders.enabled). Local automations can now be scheduled with five-field cron expressions.
For companies, the allowedMcpServers list in the managed policy now takes priority over access granted only through the MCP registry. The #installExtension, #runCommand and #newWorkspace tools are also removed, along with the /new command.
For .NET developers: if your team uses VS Code with MCP servers, check the allowed-MCP policy and whether any workflow depended on the removed commands.
Source: Visual Studio Code
Text generated by Claude Mythos 5, Opus 4.7 and Sonnet 4.6 now carries a watermark
The Claude platform release notes record on November 3 that text generated by Claude Mythos 5, Opus 4.7 and Sonnet 4.6 now carries Anthropic's text watermark. On Sonnet 4.6, responses with non-default temperature, top_p or top_k may come out without the mark.
The support article explains that the mark is imperceptible, is part of Anthropic's commitment to the code of practice under Article 50 of the European AI Act, and applies to the API, Claude, Claude Code and cloud partners. Detection is in private preview for groups such as regulators, the press, researchers and companies checking their own outputs. Anthropic warns that finding the mark is not definitive proof, and not finding it does not prove the text is human.
For .NET developers: nothing changes in your code, but if you post-process the model's output, it helps to know the mark exists and that changing the Sonnet 4.6 sampling may remove it.
Source: Claude Platform, release notes
Parallel.ForEachAsync can become a serial loop in a one-CPU container
An article on the Stack Overflow blog, by Guilherme Appel, shows a lab where reducing the container from two CPUs to one almost doubled the time to process 200 records: the median went from 42.1 to 83.6 seconds. The cause is Parallel.ForEachAsync without MaxDegreeOfParallelism set, which uses Environment.ProcessorCount as the limit. With a one-CPU limit, the loop processes one item at a time, even when the work is almost all I/O waiting.
The author's proposal is to define concurrency budgets per dependency (for example, with SemaphoreSlim), instead of a single number for everything. In his test, this approach eliminated HTTP 503 rejections in the simulated system when it was degraded. The lab's services are simulated, which the text itself makes clear.
For .NET developers: look for Parallel.ForEachAsync loops without an explicit MaxDegreeOfParallelism, especially in workers that run on Kubernetes with a low CPU limit.
Source: Stack Overflow Blog
Anthropic releases FlashProve, a prototype to prove which model generated an answer
Anthropic's Frontier Red Team published FlashProve on November 2, an open-source research prototype (Apache 2.0) for verifiable inference: a cryptographic proof (zkSNARK) that an output came from a specific model, without revealing the weights.
The numbers show it is still research. With Llama 4 Maverick on an H100, the proof is generated at about 110 tokens per second, takes close to 19 MB per request of 1,024 tokens, and costs 70 to 95 times the GPU time of running the model in vLLM. The proof covers greedy decoding and context of up to 1,024 tokens, and verification runs on CPU.
For .NET developers: nothing to adopt now; it is a topic to follow if you depend on a third-party provider and need to prove to a client which model answered.
Source: Anthropic Research
My take
For me, the news of the period is the skills on NuGet. It tackles a common problem with agents: they write code for the version of the library they remember, not the one in the project. If the instruction ships inside the package, in the right version, the agent gets the same source as the developer. The other side is that a package now also carries instructions for an agent, and that goes into the supply chain calculation along with the code.
The watermark and FlashProve answer the same question, where this text came from, but they are at very different stages: the first is already on in the models, the second is a costly prototype. The Stack Overflow piece is the most practical item in the roundup: a default that looks harmless and only shows up when the container shrinks. And .NET Conf starts on the 10th, with the .NET 11 launch, so the next roundup should be much more .NET.
Also read
Need help with .NET backend and applied AI? See the service.