And Gemini CLI github action (this project) runs again in a VM (github action runner) on a separate checkout of the code. This is what OP meant with multiple coding agents.
> This is what OP meant with multiple coding agents.
It may be the same coding agent behind the GHA. I question the implicit declaration behind OPs critique: that all 160,000+ Google folk should offer a single coding agent to their billions of users (or whatever the TAM is for coding agents). This is akin to criticizing Google Cloud for having VM, Kubernetes clusters and AppEngine; superficially, these products solve the same problem.
FWIW - this Github Actions integration is close to my ideal AI agents workflow[0]. I don't want to metaphorically look over my agents shoulder as it works in a specialized, vendor-locked IDE. I want agents to work asynchronously, taking however long they need, and tackling multiple tasks, with PRs/CLs as the unit of work. Current models may not be up to the task of single-shotting this, but the task is parallelizable across multiple agent-instances & the best solution selected (climate change be damned). I suspect Github alone may not provide adequate context as it may be missing previous tickets and design documents & the back-and-forth on requirements, but it's a start, and I'm glad Google is exploring this path for agents.
0. I believe in this workflow so much I created a proof-of-concept project that reads tickets from Vikunja and creates PRs using Aider some weeks back.
