So on one hand the page says

> We want to make dealing with agentic infrastructure easier so you can focus on your work. AX is designed with an uncompromising focus on ergonomics, rapid iteration, and joyful workflows for both application developers and AI researchers.

On the the other hand, the readme quickstart section says

> You need a Kubernetes cluster, ko (brew install ko), a container registry your cluster can pull from, and a reachable Agent Substrate Control API (in-cluster default: api.ate-system.svc.cluster.local:443).

Call me old-fashioned but I don't find this "easier". Maybe it's easier in the same way that Kubernetes itself is easier than managing VMs and container deployments at massive scale without such a tool. But there's a vast chasm between what this tool is being sold as and what it actually is.

Could someone explain to me what the general workflow is now that people are converging to? I haven't really been catching up with the AI ecosystem but I was looking into agent sandboxes and VM's recently and there's a ton of these startups and tools now. Is giving the agent a temporary scratchbox really that valuable?

I've been still just like, making VM's with proxmox, then putting my agent in the machine and letting it run free (with my dotfiles setup script making dev env pretty much free, though I could also just make a VM snapshot). What's wrong with that? Is that not the scalable solution for enterprise rn?

I have been happy with Google's Antigravity harness and Jules so looking forward to playing with this. Thanks for sharing. Simultaneously I am looking to also revisit local offline models.

While I feel like I have a decent understanding of the model landscape I'm feeling a bit lost at which agentic harness to leverage for local models. Hermes, Cline, Aider, Qwen Code, Goose, Pi, OpenCode, something else? I live in the terminal so Desktop UX is a bonus but not a must have.

Can I modify the antigravity settings/program to point to a local model? Where should I spend my energy?

The reality with releases like this is that I'm 90% sure most Google bigwigs have never heard of it, and it's misleading to label it as "Google's" in the title.

Yes, it was developed by Google employees, that does not imply it has the full backing of Google, or Deepmind, or GCP. Notably, the website doesn't seem to claim this either.

Nobody has a use for this, and anybody who can look at this website and work out what it's for is kidding themselves. Even the demo gif playing just has them pausing a task and resuming the task.

I never quite understood why agents should be treated as anything but normal software engineering. "Just" build a normal service and add an async call to some agentic framework, then parse the results. There is no need to "flip" this system and have the agent BE the process and invent a whole new ecosystem to manage the complexity that this flip creates.

If an agent is treated like nothing but a call to an external service (...which it is), everything fits in the existing programming paradigms. But I guess that's not very exciting. Only pragmatic.

> Task declares the container image and command, compute requests and limits, environment variables [...] Declares listeners the task exposes and an egress allowlist of hosts and ports the sandbox may reach. Use it to restrict an agent to, say, your LLM provider and your Git host.

I'm planning to buy a whole linux mini-PC to run my agents/code servers for more isolation. Codex/Claude Code let you run prompts on code over ssh (same with most IDEs) even on the desktop apps.

I wonder if that's going to be the new standard practice. You get a work laptop and an isolated agent box.

Running access control and network whitelists is always a maintenance challenge and it's easy to make mistakes.

> Drawing on agentic runtime research from Google DeepMind alongside deep experience in large-scale isolation, resumption, and scheduling, AX is being built as an open, declarative control plane purpose-built...

The project seems like an open-source initiative born out of the experience of some Googlers but not being used at Google. So, the title appears a bit misleading - people will be misled.

I am confused by this. The UX of the CLI is almost identical to kubectl, but it doesn't seem to actually be built on K8s CRDs. There's a backend that runs on k8s, but doesn't seem natively integrated.

Because it's using `ax apple` instead of `kubectl apply` you can't use tools like argocd for managing resources.

It's built on agent substrate which is built on top of k8s CRDs, so I'm surprised.

Is the throughput of these objects too high for etcd?

I can understand why it was chosen, but I'm not a fan of writing a bunch of yaml.

I looked at the website, and I still don't understand the purpose.

Question: What is Google's track record for where their open source releases end up over time?

Genuinely not knowledgeable here

I'm not sure why, exactly. But I don't pay any attention to news like this from Google. I don't know if there's some marketing which has me writing them off or if it's something else.

What I do know is that the Gemini integration into sheets is surprisingly incapable of performing basic tasks. This is where I expect Google to really shine. I expected Sheets + Gemini to be magical like Google Photos was. I hardly try anymore besides some basic math questions when I don't feel like inputting the formula myself.

The other thing I know is Google's propensity to sunset products. For many things, it's not a huge deal. And it may not be for this. But, why? When there are alternatives - both open and closed.

undefined

undefined

The real difficult in arranging agents is not to run them, but to identify the state change - to judge whether an agent stops to wait for you, or is stuck, or finished, and whether the two should be handled automatically or someone should be found. You have made this judgment for 12 agents, and you need to know how unreliable it is.

Everyone and their mother are vibe coding their own solutions like this, all the time.

I've been using https://github.com/mastra-ai/mastra which is pretty similar but has workflow visibility and a number of templates.

For a generic swarm, workflows aren't too useful which does away with the visibility, so I may give this a try instead.

We have built similar abstractions directly on top of Kubernetes [1]

I was looking at this project a couple of months ago, and I did not understand why not use Kubernetes instead of rebuilding the abstractions. The reason is that Kubernetes already provides other abstractions to run services and batch job, gang scheduling, gpu and other accelerators enabled workflow.

[1]: https://polyaxon.com/docs/sandboxes/overview/

Interesting, we had developed a very similar framework for our internal agents: https://github.com/apoxy-dev/clrk For us main use-case was intercepting all network I/O including LLM providers, HTTP, and random TCP/UDP calls

I am wondering how this is related to https://agent-sandbox.sigs.k8s.io/ ? Since there is also https://docs.cloud.google.com/kubernetes-engine/docs/concept...

Can someone clarify the use case for this? What's the benefit over this: https://openai.com/index/introducing-the-agents-api/

undefined

One thing I’d say, is that I find it progressively more interesting/fun to rollout your “everything”. I mean… hello security, but by the time I’ve read somebody’s documentation I’ve already halfway done making thing exactly how I want it.

> We want to make dealing with agentic infrastructure easier > Kubernetes

Pick one.

People can afford to run billions of concurrent agents?

How many months before Google kills this in favour of the next shiny thing?

Why kubernetes? Seems like an overload

I have an application usecase where this will be very helpful indeed.

I'm curious, what type of problems is this tooling aimed to solve? Isn't it a bit of an overkill regular webdev i.e.?

I really don't think any of these SOTA labs are doing agentic engineering correctly. Skills are the universal language of all agent harnesses. If you abstract the taste and prescription out of the skills and into guidance docs, then leave the skills as basically just workflow scaffolding, you can build task-specific workflows that work with any harness like Claude Code, Codex, Antigravity, etc. Technically, you only really need 2 skills, work and review, and with these you can build infinitely complex workflows including self-improving loops. I built this out and have been using it for months. It's been extremely nice. https://github.com/DanMcInerney/orchflows

Have they axed it yet?

Interesting approach. How does it compare to LangGraph for multi-step agent workflows? The orchestration layer always seems to be the hardest part to get right in practice.

undefined

How does it compare to kagent (https://kagent.dev/)?

So the agent-substrate checks a _ton_ of boxes. Almost all of the things it offers should be table stakes for everywhere we run not only agents but most software.

https://github.com/agent-substrate/substrate

(For context I built something very similar to this the past 2 weeks for my homelab, trying to solve many of these problems. This comment is an edited version of an unreleased blog post I wrote last week.)

- Run code in secure microVMs or gVisor. Docker is not good enough. Qemu is not good enough. A secure environment for running untrusted code is the bare minimum. I don't see Firecracker in the repo yet, but that's ok the idea is there.

- Fast resumption. In my homelab, time-to-first-message is around 11-12 seconds. That's half setting up the pod, and half resuming the CLI (e.g. `codex resume ..`). Why resuming? In my homelab agents are commonly blocked waiting for CI or waiting for me to approve an action, in this case I stop their container to keep resource usage low. Then for resumption, you definitely don't want to waste the agents time by giving a new ephemeral disk and forcing them to re-clone and re-build. For microVMs this is not actually straightforward, for example Firecracker only allows block devices, so re-attaching an agents disk workspace requires a custom storage interface

- Zero Trust. Codex CLI permissions for example are extremely broken. "Can I run this 500 line long command? or allow any command starting with first 100 chars always?" More reasonable grants are needed.

I don't understand yet how they will surface Zero Trust notifications. In my homelab it's a Forgejo comment linking to an auth service, and a ntfy.sh iOS notification which opens up the auth service.

I don't get why they to restore the RAM of the agent env. Maybe to fully optimize resumption. Idk, I don't have that much RAM in my homelab, my agents use a ton, testing stuff in Chromium making screenshots for me. I can't keep RAM for 100 workspaces from the past 24 hours in RAM.

MITM gateway is very cool.

I'm curious how they will integrate with microVMs. I just wrote yesterday[1] about how there are NO GOOD OPTIONS for this atm. Kata is decent but the attack surface it introduces makes me uncomfortable.

[1]: https://srcreigh.ca/posts/auditable-kata/

But anyway, even if this project is abandoned out of the gate by Google, we should be happy, it sets the bar where it should be. I'm excited to learn how they solved these problems differently than I did.

I'd evaluated both Google's Agent Substrate (that underlies Ax) and their Scion project. I really enjoy how Scion operates with existing tools really well. Ax/Agent Substrate is much more a greenfield independent effort, it's own thing.

I think Scion has so much more mature a disosition: you could write OpenCode plugins that enhance the runner, and use that locally, and use it in Scion. With Ax/Agent Substrate, you are opting in to a pretty huge stack that is just Agent Substrate, that is their runners, their harness, their substrate. I do think their actor model is pretty neat! It's neat having the agent have such primacy! But it feels so much less integrative, is such it's own thing. Scion, to me, is much more interesting an effort, that similarly helps scale out agentic workloads.

https://github.com/googlecloudplatform/scion

undefined

Oh no, Kubernetes for agents. I guess all roads lead to complex YAML.

Anyone who uses this promotion packet fodder for anything important is a fool

Is the logo a cheerful little parasitic skin mite?

Surprising no mention of Google on the page or domain.

Kubernetes but for agents :(

I don't see a meaningful difference to the 100s of other 'agentic frameworks' that promise to be the one to all solution for all your troubles.

Would be about time we get benchmarks for these ... so these can also be gamified just like with the LLMs.

Yeah, my grandmother also did something on this https://github.com/nillebco/varda

Essentially there is no out of the box solution about orchestrating agents and increasing LLMs sandboxing. That's why everyone and their grandmother are re-inventing the wheel.

At the same time, it's an incredibly complicated problem, with a variable perimeter (OS support, sandboxing primitives support).

I am quite happy about my own solution (because it supports my use case!) but I hope something with a decent dev UX will appear one day. AX definitely is NOT.

The terminal gif is the most confusing slop I've seen from Google. It doesn't explain anything and it just seems to be a collection of random commands that someone ran to test, not something that tries to explain what the tool does.

undefined

This is bound to cause some confusion with the other tool called Ax for agentic development: https://axllm.dev/ (which is DSPy for other languages)

I was expecting that, in the AI era, even Google will start using Rust for everything. But they chose Go for the this project.

How is this different than k8s jobs?

They will sunset this in 6 months. Dont bother

Kubernetes but for agent ?

It's crazy how far behind Google has fallen in this space in just a single year

It looks great from a security standpoint, but it also feels like overkill.

not exactly for agent orchestrator, but I did make something similar in terms of design: https://github.com/yash-srivastava19/cadence

AX, not to be confused with Ax the machine learning tool from Meta for optimizing experiments.

https://ax.dev

k8sification of AI was always inevitable, if only as a form of salary justification.

I'm keeping an eye on another Google Cloud orchestrator

https://googlecloudplatform.github.io/scion/overview/

Scion wraps the harnesses (9x) we all use every day and is closer to OpenClaw on Kubernetes

Since hearing the word orchestrator in ai context it was clear that ClanKernetes was coming.

this is nice, basically virtual threads for kubernetes.

This is different from langchain etc because lanchain works at the app layer but this one works at the infra layer with tool calls etc?

Unnecessary complexity packaged as product!

Am I being gaslighted into thinking over engineered systems are not?

Kubernetes is the last thing I wanted to see recreated for agents. It’s like Multics of cloud, now for agents. Complexity for the sake of it, powered by your favourite YAML slop bowl.

As usual, Google makes it "googley" by building an incompatible monolith with the kitchen sink included.

I just have a tmux session acting as the orchestrator, and I tell it to report back and direct the other agents working in separate tmux sessions.

"2. Deploy the control plane

You need a Kubernetes cluster"

LOL. Bye!

A DAG?! Holy innovation, Batman!

[flagged]

undefined

[flagged]

undefined

[flagged]

[flagged]

[flagged]

[dead]

[dead]

[flagged]

[flagged]

[flagged]

[dead]

[dead]

[dead]

[dead]