I have been building an OSS self-hostable agent infra suite at https://ash-cloud.ai

Happy to trade notes sometime!

On write amplification — we persist at the message level, not per SSE chunk. The sandbox's workspace filesystem (claude code's native jsonl files) is the source of truth for resume, and the DB is for queryability, tracing, etc - so fire and forget works fine here.

I kept rebuilding the same infrastructure every time I wanted to put a Claude Agent SDK into production. Sessions, streaming, sandboxing, persistence, a REST API, file hooks. So I built Ash.

An agent is a folder with a CLAUDE.md file, skills folder, etc, and one-line deploy from cli, or tsx/python sdks.

Some things I cared about when building this:

Sandboxing. Each agent runs in its own isolated process with an environment allowlist, cgroups resource limits, and bubblewrap filesystem isolation on Linux.

Session persistence. All state lives in CRDB. If the server crashes mid-conversation, resume picks up where it left off. You can also snapshot workspaces to S3/GCS and resume on a different machine.

Keeping it thin. The SDK's `Message` types flow through the whole pipeline untranslated, from the Unix socket to SSE to the client. Ash adds session routing, sandbox pooling, and lifecycle management. It doesn't re-wrap or translate the SDK's types.

I've been measuring overhead pretty carefully. Ash adds <0.5ms per message at p99. Warm resume is 1.7ms. Cold resume (restore workspace, spawn process, reconnect) is 32ms.

OSS, Self-hostable, MIT licensed: https://github.com/ash-ai-org/ash-ai

Comments URL: https://news.ycombinator.com/item?id=47291795

Points: 1

# Comments: 2

With an agent I know if I write once to CLAUDE.md and it will be read by 1000’s of agents in a week.

Can we replace the link with the actual blog post that doesn't include/promote KaraVideo? https://openai.com/blog/sora-first-impressions

These sort of devices are nuisances with very low positive utility, and there is plenty of precedence for banning them.

TFA argues that digitizing one self inevitably leads to two copies, and the digital one is not real.

Let’s take a step back and start by replacing/simulating one neuron digitally. Is this a new you? I doubt TFA would argue so. Now replace 3,000 neurons every second over the course of a year and now your entire brain has been digitized. At which point did you stop being you?

The options are, 1) every second you are a new you, with or without digital replacement, 2) every second you are the same you with or without digital replacement 3) every second you are a new you with digital replacement but the same you if you don’t digitally replace. 4) some arbitrary % that feels right like 50%.

Option 3 hinges on a special differentiation between digital and chemical computation, which is at odds with fundamental properties of Turing Machines. Option 4 is as hand-wavey as it gets. Options 1+2 while they seem very different, ultimately can be treated as functional equivalents with slight philosophical differences.

There are real debates to be had over the morality of mind uploading and the very real risks of doing so, and many great sci-fi short stories covering this topic. Unfortunately, this article is too caught up on its hot take to make room for a proper consideration beyond its knee-jerk reaction.

Side note: I worked with Kyle a few years ago on the MIT-MGH Deep Learning for Mammography project. I'm glad to see his work + brilliance being recognized.

1. CNN's are used fairly commonly for sequence tasks nowadays. Convolutions can be 1D after all. 2. It's also possible the previous comment was referring to using 2D convolutions on the spectrogram of the audio, which is a common approach. 3. Neural networks are capable of more than classification. Scoring is a regression task which is common application of neural networks.