Google's AX agent runtime hits v0.3.0, scaling to millions of tasks
Apache-2.0 and written in Go, it drops Kubernetes CRDs for Redis Streams so etcd stops capping agent scale — and CodeMidas turns repos into RL tasks.

Copy markdown
Three services, and Redis instead of etcd
v0.3.0 breaks AX into ax-server, ax-controller, and ax-task-runner and moves task state out of Kubernetes custom resources into Redis Streams. That single change is what lets one cluster run millions of short-lived agent tasks without hammering etcd into the ground.
You describe an agent in four YAML objects
Everything is expressed with just Task, Workspace, Gateway, and Model — the sandbox, the git repo, the network policy, and the model credentials. If you've ever written a Kubernetes manifest you already know the shape; `go install github.com/google/ax/cmd/ax@latest` and you're running.
The catch: this wants real infrastructure
AX is not a laptop toy. It expects a Kubernetes cluster, a Redis instance, a container registry, and LLM API credentials before it does anything at all. This is aimed at teams putting agents into production, not weekend tinkering on a single box.
Build this weekend: a fleet of throwaway agents
Because state now lives in Redis Streams, spinning up thousands of one-shot agents — CI fixers, PR reviewers, scrapers — no longer melts your control plane. Point AX at a queue and every task gets its own sandboxed, resumable runner that cleans up after itself.
Elsewhere: CodeMidas turns any repo into RL
A new paper builds executable RL environments straight from source code — 5,545 tasks across 3,185 codebases in 23 languages. Training MiMo-V2.5 on them lifted ProgramBench +17% and DeepSWE +11.7%, a hint that your own repos are untapped agent-training data.