gianni

Building @nightshiftdotsh

[gianni @tensor_ninja]

@nightshiftdotsh is a young project with ambitious goals. We want to be the serverless runtime for AI agents and provide an incredible dev and operator experience.

We're not yet, but with the help of the community we'll get there. Is and will always be open source.

We're building software on github https://t.co/CctqnmRjpO

and building community on slack https://t.co/BGc3ZGN28N

[Nightshift @nightshiftdotsh]

Deploy Claude Code SDK Stateful Agents into isolated sandboxes that run in the cloud but feel local.

Build your flock of Agents 🦜

[Armin Ronacher ⇌ @mitsuhiko]

I think it would be great if people were upfront about declaring their own understanding of a topic / their pull request. Now that everybody can talk confident with their clanker it becomes way too hard to understand if they knew what they were doing when they prompted it :(

[NVIDIA AI @NVIDIAAI]

@xeophon @arcee_ai Open > closed

[gianni @tensor_ninja]

@a16z https://t.co/y3t8AtUnTj is Sovereign AI.

[gianni @tensor_ninja]

$150k of angle funding to get us here https://t.co/y3t8AtUnTj

An agent harness that scales.

[ethan @henl_ey]

lots of discourse on context and not nearly enough on making search + access better/more scalable for agents

[gianni @tensor_ninja]

Is there a serverless type platform where I can bring a LoRA adpater and inference is just handled for me?

[gianni @tensor_ninja]

@__tinygrad__ @APompliano yaaaas queen 👏

[gianni @tensor_ninja]

why does docker desktop suck so bad

[gianni @tensor_ninja]

@erikdunteman The agent should be in a sandbox. The code the agent produces should be in a sandbox. The actions the agent takes should be in a sandbox.

It's just so much more productive. I will die on this hill.

[gianni @tensor_ninja]

@aniketmaurya Oh that’s so hot

[gianni @tensor_ninja]

Our current runtime is held together with duct tape and bubblegum. By that of course I mean Python asyncio and sheer determination.

We're starting the process of embedding containerd directly into Nightshift to standardize the communication between the Agent API Runtime (nightshiftd) and the sandbox.

The goal is to have a flexible platform that can run agents in any sandbox you want (Kata Containers, Firecracker, runc, whatever) across any cloud or on-prem, inside or out of Kubernetes.

I'm documenting this journey and as our first installment. I wrote up how I wrapped my head around container networking.

https://t.co/nmeo5Z1aYv

[gianni @tensor_ninja]

@Gavriel_Cohen Yup. Thin is good people.

[gianni @tensor_ninja]

@itsandrewgao You’d be surprised how much further you can push it when you just know what things are called…even in system engineering.

[gianni @tensor_ninja]

Here is the code https://t.co/k7lAOwEVYa

[gianni @tensor_ninja]

I was tinkering with KVM to get an idea of what Firecracker was doing under the hood for @nightshiftdotsh and decided to write a blog post of building a VMM (what firecracker is) in Python, from scratch.

https://t.co/LqmME8fHDc

[gianni @tensor_ninja]

In the latest release of @nightshiftdotsh (v0.4.3) Agent events are durable and repayable.

Deploy your own KVM backed Agent Sandbox Platform. Is and will always be OSS.

[gianni @tensor_ninja]

The @nightshiftdotsh SDK is boring.

We designed it that way.

Write your agent in the frameworks you're used to (we're using the @OpenAI Agent SDK here). Wrap your agent in a decorator. Run it in a sandbox. Deploy and serve users.

Is and will always be open source, for the community.

[gianni @tensor_ninja]

@brandonturp We're building @nightshiftdotsh in Miami open source with the community! Would love to be apart of your events.

[gianni @tensor_ninja]

@wildmindai Was literally in the middle of building a script that does this. This is much better. I love OSS