"How Do You Fix 70% Data Loss Across 1 Million Concurrent Connections?" - this is a question a Production @AkkaDotNET Support customer asked us earlier this year. We solved it.
Here's how we did it:
A thing I haven't talked much about yet: Netclaw ships an open source "skill server" that exposes the proposed @Cloudflare agent skills standard for being able to discover / version skills just like a package feed essentially.
Idea is to make it trivial to sync skills across org
New video / blog post: "Why We Removed FluentAssertions from @AkkaDotNET "
Akka .NET 1.5.64 removed FluentAssertions from every TestKit package.
FA went commercial, and we were the 2nd-largest consumer of it on NuGet. No shade. But we won't hand a commercial dependency to every
Using SkiaSharp to resize / compress images for multi-modal LLMs in Netclaw github.com/netclaw-dev/ne…
Should make uploads more token and bandwidth-efficient
Netclaw v0.23.0 is out - lots of bug fixes and improvements in this release, including public beta build feed access:
1. Fixed numerous Docker issues (i.e. zombie processes)
2. High memory usage in tool-heavy call sessions fixed
3. Fixed Codex HTTP compat issues
4. More!
If Incremenalist suddenly broke for you, it's because changes to the most recent .NET 10 SDKs moved some MSBuild bits around. Several MSFT-hosted AzDo images all got upgraded to use that SDK.
Update to Incrementalist 1.2.1 and that'll fix it. Please RT
Netclaw v0.22.0 is out and includes a ton of bug fixes and improvements that allow it to handle larger, longer-running jobs including those triggered by webhooks or scheduled reminders.
👇
So if you're just hearing about Netclaw for the first time and wonder what the hell Aaron is talking about, let's introduce you to "prefix caching" - this is what makes long context windows feel responsive & fast when working with inference providers for large language models 1/5
Can't easily do this as part of CI/CD due to just the sheer instrumentation overhead, but when working on Netclaw's context construction pipeline you always need to benchmark your prefix K/V cache re-use with inference providers
Agentic harnesses can BUST the cache by inserting something anywhere other than the tail end of the context - this invalidates everything underneath it, which requires full re-processing for that portion of the prompt. This is slower and possibly even more $ expensive.
Netclaw (.NET agents) v0.20.0 is out!
You can now use your @github CoPilot subscription as an inference provider.
You can now use @Mattermost as a communication channel.
Reverse-proxy is now a first class exposure mode.
And lots and other bug fixes and improvements. 1/3