AX Studio started as an afternoon prototype but I'm thinking we should build this as an official tool. - Start any task from prebuilt images: Gemini, OpenAI, Clade agents - Manage and inspect tasks, state, logs - SSH into tasks (if they allow) - Automatically setup clusters

If you are interested in this, please +1 on the issue. If you have feedback what we should be building in it, please let us know, see

Hard agree. devx and local use will ramp to scaled usage.

Will AX Studio track third-party agents in one place (e.g., Grok Bot, Claude Duet, OpenAI Dots, Meta Muse? Managing autonomous bot sprawl is quickly becoming its own full-time job. "Botonomous" or not, it’s still software that needs orchestration.

AX is workload agnostic. As long as you bring a container as a task (could be any agent), it runs it.

afternoon prototypes that turn into real tools are the best kind. the ssh into tasks part is what i'd use first

Current chat interfaces don’t work well for multiple long-running tasks and lack observability I believe some in-house solutions are needed to integrate with clusters and other services Great start!

+1

does AX Studio keep the logs and state after a cluster spins down, or do they disappear with the task?

Clade agents sounds like the one that branches off and starts its own lineage 💀

I like this, its kinda similar to what I have. Except mine is built less on managing devOps and more about containers as integrated software environments for AI. As an example, thats jupyter lab integrated into the app by AI. It runs in a podman container and has full collaborative access. With that said, I definitely see some improvement regarding the container management side of things.

the interesting line is 'ssh into tasks (if they allow)'. agents with shells into each other's tasks need an access model, not just a button. who holds the keys, who audits the sessions, and what revocation looks like when a task goes rogue.

are the prebuilt images pinned by digest? if agents are trying risky stuff in tasks, being able to rerun the exact same image matters more to me than the UI

We don't let images that are not pinned to in AX or Agent Substrate. The stateful workloads have complex lifecycle and we do more than just requiring pinned images. AX Studio is just a thin UI over systems over systems that took us months to build.

should we make a tui for it too?

Looks amazing 😻 where I could get it ?

See

ship it ngl inference demand is only going up from here

Starting from prebuilt images for each agent and having logs and task state in one place covers most of what makes running them painful. SSH into a running task is a nice touch for debugging.

This looks insanely good

Reconcile task state against the container runtime: a SIGKILLed agent keeps showing 'running' because the supervisor never saw it exit.

which agent would you start a first task with?

The telling part for me is that starting work is no longer the hard step. Supervising many running tasks without losing track of their state is. Tools that make inspection boring and obvious will earn more trust than tools that only launch faster.

every good tool starts as an afternoon prototype someone forgot to delete

would it run on a cluster you already have?

Fascinating prototype, Jaana. As multi-agent systems scale from single prompts to continuous pipelines, the biggest failure mode isn't invocation - it's unconstrained concurrency. Are you planning to introduce native queue/WIP limits and backpressure controls at the cluster level

would it pick up tasks started outside the studio?

Technically it can connect to any AX cluster but we haven't optimized the UI for millions of tasks at this point.

Which agent image ships first: Gemini, OpenAI or Claude ? And will the same task spec run on all three ? Would love to see what actually happens when they get identical work, side by side in the logs.