Google's AX Runtime Resurges on GitHub, Reliability Debate Unresolved

Google's open-source agent orchestrator, AX, sat at the top of Hacker News for most of Sept. 20 and 21, drawing 662 points and 299 comments on a post titled "AX – Google's Open Agentic Orchestrator." On Sept. 25, the google/ax repository picked up another 1,373 stars in a single day, ranking second among GitHub's daily trending repositories, behind only an agent-memory tool called Hindsight.

None of that is a launch. AX shipped as a preview four months earlier, on May 20, when Google software engineer Jaana Dogan and engineering director Ethan Bao introduced it on the Google Cloud blog as Agent Executor, "an open-source runtime standard for agent execution, resumption, and distributed deployment." What changed between May and September is attention, not code: the repository still carries a warning that its "core concepts, protocols, and specifications" are actively being refined and that breaking changes are likely before a stable release.

What Agent Executor actually does

AX runs on top of a separate Google project called Agent Substrate, which Dogan and Bao describe as an agent-first abstraction layer over Kubernetes, built to handle "millions of sub-second tool calls" and scale toward "hundreds of millions of registered agents." AX itself is the orchestration layer on top. It sandboxes an agent's task, wires up its workspace, and exposes a deliberately kubectl-shaped interface: apply, get, describe, watch, delete, plus a handful of agent-specific verbs for suspending and resuming work.

The pitch is durable execution. An agent that has been running for hours, waiting on a human approval or a slow tool call, can be checkpointed and resumed rather than restarted from zero. Sessions stay consistent under concurrent writes, disconnected clients can reconnect and get a backfilled response, and a task's history can branch from a saved checkpoint to test an alternate path. Dogan and Bao frame this as the gap between a weekend demo and a system a company can leave running: a demo agent that crashes gets restarted by a developer watching a terminal, while a production agent that crashes mid-task, with a half-sent email or a half-written file, needs a record of exactly where it stopped.

The project is Apache 2.0 licensed, written mostly in Go, and as of Sept. 25 carries 10,441 stars, 505 forks and 38 open issues on GitHub, a repository created on March 30.

Broadcom and Avasant aren't sold

The closest thing to independent scrutiny AX has received came five days after the May preview, when InfoWorld senior writer Anirban Ghoshal quoted two people outside Google on May 25. Advait Patel, senior site reliability engineer at Broadcom, agreed with the diagnosis: "Durability, orchestration, and resumability are the real blockers for any enterprise production agents," he told InfoWorld, adding that frameworks like LangChain and AutoGen "more often than not fall apart in production once agents run for hours or days."

Gaurav Dewan, research director at Avasant, was less convinced the runtime itself is the fix. A scheduler that survives a pod restart doesn't answer for "accountability, explainability of agent decisions, policy enforcement," he said: the governance questions that show up once an agent is making decisions in a real system, not just staying alive while it makes them. Those are exactly the questions a Kubernetes-shaped tool is built to be silent on, since Kubernetes itself has never tried to answer who is allowed to approve a deployment, only whether the deployment stays running.

The number on GitHub today

AX's Sept. 25 trending spike sits inside a broader pattern. Five of the day's ten most-starred GitHub repositories are agent infrastructure rather than agents themselves: memory layers, orchestrators and harnesses that wrap around whichever model a team is already using, rather than new models or new agents that use them.

1. vectorize-io/hindsight1668 stars
2. google/ax1373 stars
3. dream-num/univer1082 stars
4. obra/superpowers611 stars

Source: GitHub Trending, accessed 2026-09-25

google/ax on GitHub	Sept. 25, 2026
Stars	10,441
Forks	505
Open issues	38
License	Apache 2.0
Repository created	March 30, 2026

The repository's own warning about breaking changes hasn't stopped the star count from climbing roughly 15% in a single day, off the strength of a four-month-old blog post recirculating on Hacker News rather than any new release. That says something about how the agent-tooling category gets attention now: one loud thread can outdraw a formal launch. It is not evidence the reliability problem Dogan and Bao describe is actually solved.

What isn't public yet is any account of a team running AX against production traffic rather than a demo. Google's own post names no customer, and neither InfoWorld's reporting nor the Hacker News thread surfaced one in the four months since the preview shipped. Until one does, or until Dewan's governance questions get an answer that isn't left for a future release, AX's Sept. 25 numbers describe interest in a fix, not proof that the fix works at the scale Google is claiming.

GitHub-trending agent tooling has produced overclaimed projects before, and AX's own vendor-supplied capacity figures, "hundreds of millions of registered agents," have not been independently tested at that scale by anyone outside Google. A narrower, already-shipping piece of the same puzzle is agent memory, which AX does not attempt to solve. Dogan and Bao's post treats session state and long-term memory as adjacent problems, ones a team pairing AX with a separate memory layer would still have to solve on its own.