Google open sources AX, an Apache 2.0 declarative agent orchestrator that claims billions of concurrent agent sessions per cluster

I skip news like this from Google now. Gemini in Sheets can't do basic stuff I'd write a formula for, and Google sunsets products all the time. Why bother when there are open and closed alternatives?

It's Apache 2.0 and the code is on GitHub, so the sunset argument matters less here.

I gave the Gemini integration in Docs and Sheets a real shot and it's useless. I asked it to convert my Markdown markers into Docs formatting, two pages, just headings, bold, italics, and code blocks, and it destroyed the formatting through half the doc. GPT 3.5 could do that.

Same in Gemini notebooks. I dropped my CV in and asked for a cover letter for a job description, then three messages later 'rewrite that friendlier' produced a letter for a different job from another chat in the same notebook.

AI getting k8sified was inevitable, if only as salary justification.

Genuine question: what's Google's track record for where their open source projects end up?

Maintain it for a bit, then let it die slowly.

They forked their own Android project to patch security holes and never upstreamed the fixes, so that's one data point.

Depends. Kubernetes, Go, TensorFlow, Chromium, gRPC all went incredibly well.

I get why they picked YAML but I'm not a fan of writing a pile of it.

I evaluated Agent Substrate, which Ax sits on, and Scion. Scion plays well with existing tools: you can write OpenCode plugins and use them locally and in Scion. Ax makes you opt into their runners, harness, and substrate as one big stack. Their actor model is neat and giving the agent primacy is neat, but it's much less integrative.

The site says the opposite. It's described as low opinion and the running tasks don't even have to be AI agents. How are you more locked in than that? Scion talks way more about concrete agents.

Most Google bigwigs have never heard of this, so putting 'Google's' in the title is misleading. Being built by Google employees doesn't mean Deepmind or GCP backs it, and the site doesn't claim that either.

The repo description at github.com/google/ax says 'Google's open agentic orchestrator', and it's under Google's GitHub org.

rakyll is working on it and tweeting about it, there's a real team behind it, looks like a GCP effort.

There's a formal announcement on the Google Cloud blog from 4 months ago.

This will confuse people with the other Ax for agentic work, axllm.dev, which is basically DSPy for other languages.

I don't see a real difference from the hundreds of other agentic frameworks that promise to solve everything. We need benchmarks for these so they can get gamified like the LLMs.

I don't think the SOTA labs are doing agentic engineering right. Skills are the universal language of harnesses: keep the skills as workflow scaffolding, move the taste into guidance docs, and two skills (work and review) build any workflow, including self-improving loops. I've run it for months.

How does that relate to Ax? Its concepts doc mentions skill registries.