Google's Agent Executor Passes 7,600 Stars After v0.3.0

Google's open-source Agent Executor, a runtime for running AI agents as resumable, sandboxed workloads, has passed 7,606 stars on GitHub, according to the repository's own count, after its third tagged release, v0.3.0, shipped Sept. 20 with an empty release-notes field and no accompanying blog post.

The repository was created March 30 under Google's own GitHub organization and describes itself as "Google's open agentic orchestration runtime." Google software engineer Jaana Dogan and engineering director Ethan Bao introduced it on the Google Cloud blog on May 20, the same day the project's first tagged release, v0.1.0, went out. Their post argues that long-running agent workflows are "fragile and incredibly hard to manage reliably and efficiently in production" because agents are stateful and bursty rather than the short, stateless requests most infrastructure is built to serve.

Four primitives, one license

The project, licensed under Apache 2.0 and written in Go, is built around four declarative primitives: Task, an isolated sandbox execution; Workspace, a set of pre-wired resources; Gateway, a network allowlist; and Model, an LLM configuration. Durable execution lets a task resume from an event log after an outage or a human-in-the-loop pause, and Dogan and Bao's post says a companion system, Agent Substrate, is designed to register hundreds of millions of agents and handle millions of sub-second tool calls. Agent Substrate itself, along with a Managed Agents API and a tool called Antigravity 2.0 that Google also announced, remain closed rather than open-sourced alongside Agent Executor.

Version	Date
v0.1.0	May 20, 2026
v0.2.0	Jul. 21, 2026
v0.2.3	Aug. 13, 2026
v0.3.0	Sept. 20, 2026

What the release notes didn't say

GitHub's own record shows all five tagged releases were published by Dogan's account, rakyll, and v0.3.0's release page carries no changelog text at all, a contrast with v0.1.0's announcement in May, which described the project as an early-stage release meant "to be able to have some concrete conversations." The repository has gained 355 forks and carries 31 open issues as of this week, a smaller haul than the 233,000 stars DeepSeek's own open-source agent harness has drawn, though Agent Executor is infrastructure a harness would run on top of rather than a competing product. InfoQ's Olimpiu Pop, writing Sept. 22, reported that infrastructure engineers describe cost savings from eliminating idle compute, while other developers complain about "the heavy operational overhead of maintaining Kubernetes clusters, container registries, and custom CRDs" that Agent Executor's Kubernetes-shaped design still requires. Security-minded users welcomed the project's use of gVisor for sandbox isolation but flagged early problems with dropped connections through its egress proxy, according to the same report.

The idle-compute number nobody outside Google can check

Dogan and Bao's blog post does not publish a figure for how much of an agent's runtime is actually spent idle, waiting on a model response, a tool call or a human review, only that it is the majority of it. That is the same kind of unverified number TypeSafe made about its own Jev model's speed and cost claims: a vendor's own account of a problem, offered as the justification for a new piece of infrastructure, that nobody outside the company has measured independently. Pop's InfoQ report positions Agent Executor as a foundational execution runtime rather than a higher-level framework like LangGraph or CrewAI, aimed at enterprises running agent fleets rather than individual developers, which is also why the operational overhead complaints Pop documented are the ones worth watching: a runtime aimed at fleets has to prove it is worth the Kubernetes tax before those fleets show up. Google has not said how many organizations are running Agent Executor in production, as opposed to the 7,606 accounts that have starred it on GitHub.