Google open-sources AX, a Kubernetes-style runtime for agent swarms

The Apache-2.0 orchestrator topped HN but drew skepticism; also: AMD cuts KV-cache 98%, and constrained agent loops hit 32x Rust speedups.

Nowline SEP 23 1:00 AM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • What AX actually is

    AX is a Go, Apache-2.0 orchestrator that runs autonomous agents like Kubernetes pods: Task and Workspace primitives hand each agent an isolated, allowlist-egress sandbox with suspend/resume checkpointing and SSH-in debugging. If you're already on a cluster, `go install github.com/google/ax/cmd/ax@latest` gets you a kubectl-style CLI for agent workloads.

  • The catch: Hacker News isn't buying 'billions'

    Google pitches AX as running 'billions of autonomous agent workloads,' and the top HN thread pushed back hard on both the number and the setup cost — you need a Kubernetes cluster, the `ko` builder, and Google's Agent Substrate Control API before anything runs. Treat it as production infrastructure, not a weekend project.

  • AMD: 98% smaller KV-cache, 32x the context

    AMD's Zebra-HyLo post-training recipe converts stock Transformers into MLA + Mamba-2/DeltaNet hybrids, cutting KV-cache 92-98% and stretching context up to 32x (8K to 64K) for only ~10B tokens of tuning. Fourteen Apache-2.0 checkpoints (Llama-3.2-1B/3B, Qwen3-1.7B) are on Hugging Face and the recipe runs on your own weights — long context without the memory bill.

  • Build this weekend: an optimizer that can't cheat

    Max Woolf shows that handing a coding agent a concrete speed target, anti-cheating rules, and a `criterion` benchmark compounds into 7.5x-32x faster Rust across model generations — genuine algorithmic wins, not gamed metrics. Point the same loop at your hottest function and let it grind toward a number.

  • Elsewhere: Claude open-sources its model-speedup code

    Anthropic had Claude autonomously optimize 30+ open biomolecular models for an average 4x speedup, then open-sourced the optimization code plus a new low-memory inference mode that fits 10k+ token systems on a single GPU node. There's also a $1M Claude-credit competition with Adaptyv Bio for protein-design validation.