I am wondering how this is related to <a href="https://agent-sandbox.sigs.k8s.io/" rel="nofollow">https://agent-sandbox.sigs.k8s.io/ ?
Since there is also <a href="https://docs.cloud.google.com/kubernetes-engine/docs/concepts/machine-learning/agent-sandbox" rel="nofollow">https://docs.cloud.google.com/kubernetes-engine/docs/concept...
IIUC agent-sandbox and agent-substrate (the one ax builds upon) are similar. Agent-sandbox is more k8s-native, while agent-substrate is less so.
Personally I think that this kind of workload is better off not being tied too much into kubernetes. I've worked with crossplane and other controller who put a lot of load on the k8s-apiserver and etcd and can easily slow the whole machinery down / grind them to a halt.
btw, agent-substrate is in the process of being moved to CNCF: <a href="https://github.com/cncf/sandbox/issues/523" rel="nofollow">https://github.com/cncf/sandbox/issues/523
Agent Substrate was built to provide a few (important) things over Agent Sandbox:
* More efficient usage of compute by timeslicing agents (Substrate Actors), which requires fast suspend and resume (using gVisor or cloud-hypervisor snapshots), as well as keeping the K8s control plane out of the critical path (so agents can't be stored as resources in the K8s database).
* Deep inspection of outgoing requests using an egress gateway
* Minimizing the exposure of credentials to unpredictable agent control (so they can't upload access tokens to pastebin).
Achieving those goals ultimately required a significantly different design from Agent Sandbox.
je42 · · focus · HN ↗
jcw90210 · · focus · HN ↗
Personally I think that this kind of workload is better off not being tied too much into kubernetes. I've worked with crossplane and other controller who put a lot of load on the k8s-apiserver and etcd and can easily slow the whole machinery down / grind them to a halt.
btw, agent-substrate is in the process of being moved to CNCF: <a href="https://github.com/cncf/sandbox/issues/523" rel="nofollow">https://github.com/cncf/sandbox/issues/523
ahmedtd · · focus · HN ↗
* More efficient usage of compute by timeslicing agents (Substrate Actors), which requires fast suspend and resume (using gVisor or cloud-hypervisor snapshots), as well as keeping the K8s control plane out of the critical path (so agents can't be stored as resources in the K8s database).
* Deep inspection of outgoing requests using an egress gateway
* Minimizing the exposure of credentials to unpredictable agent control (so they can't upload access tokens to pastebin).
Achieving those goals ultimately required a significantly different design from Agent Sandbox.