We've talked about this a lot so let me try to write this down a bit.
high level idea
The idea is to be able to use a harness (eg agy) together with SRE skills to be able to investigate a real-world scenario.
Let me try to write down what i envision here:
Whats in the docker container
- git pull SRE ext
gcloud- some key to allow the gcloud to authenticate as SafeSRE ServiceAccount identity. gcloud will work out of the box with this identiy which is NOT allowed any mutation in project
PROJECT_ID (with a few safe exception, like "Create a Cloud Ops dashboard".
- fuse mount of GCS bucket for persisted memory. Nit: this could be TWO buckets, one pvt for investigation, one public for publishing findings and comfort of PII-free permalink you can give to anyone, but this part I havent thought through yet.
- Optional (since 2026 is the year of openclaw): a Telegram channel.
This is the engine which can do an investigation. Note that SRE extension is called as skills, and is pulled as latest build (nit: might be latest official TAG).
Rules of Engagement
- This should be triggered by an external event: an email, a message, a pub sub notification.
- Agent lives in Cloud run or Gemini Managed Agents [2].
- Investigation "spins up" this container (whatever spin up means, could be a dormient CRun or simply spin up a NEW GMA instance).
- You give it a problem to solve (an "investigation"): a broken SLO, a failed GHA push to repo X, an email or a generic prompt message "People report GKE cluster down".
- system starts to investigate and come back with a proposed solution. This might be a "do command X which i cant run but you can, and this is the risk factor" and provide additional graphs, dasboards, ... to suffragate this hypothesis.
- [OPTIONAL] there could be an "unsafe gcloud" MCP functionality which can be triggered only by phyisical user after user approves the YES, but this shouldn't be part of v1 even if my friends are already pitching it :) [1]
The WHY
Why is this cool? Well, it demonstrates an offline behaviour for our Ops agent that customers and the market find highly desireable at the moment. More context b/529228771
[1] https://x.com/jackwoth98/status/2070943018991919544?s=46&t=coFmTFsfL9HQlVP_uQJzgg
[2] https://blog.google/innovation-and-ai/technology/developers-tools/managed-agents-gemini-api/
We've talked about this a lot so let me try to write this down a bit.
high level idea
The idea is to be able to use a harness (eg
agy) together with SRE skills to be able to investigate a real-world scenario.Let me try to write down what i envision here:
Whats in the docker container
gcloudPROJECT_ID(with a few safe exception, like "Create a Cloud Ops dashboard".This is the engine which can do an investigation. Note that SRE extension is called as skills, and is pulled as latest build (nit: might be latest official TAG).
Rules of Engagement
The WHY
Why is this cool? Well, it demonstrates an offline behaviour for our Ops agent that customers and the market find highly desireable at the moment. More context b/529228771
[1] https://x.com/jackwoth98/status/2070943018991919544?s=46&t=coFmTFsfL9HQlVP_uQJzgg
[2] https://blog.google/innovation-and-ai/technology/developers-tools/managed-agents-gemini-api/