Create a technical analysis of Kagent and Kagenti and form an opinion on which one to invest in.

[usize]

Kagenti and Kagent are both Kubernetes native frameworks that aim to manage the lifecycle of Agents. They broadly achieve this through defining custom resources and operators, and by relying on the conventions of the A2A protocol as integration points. However, their implementations do diverge substantially in terms of particulars.

Architecturally, their philosophies seem to diverge primarily around how lifecycle is managed.

• In Kagent a controller maintains a graph of agent dependencies via CRDs + a single controller. Dependencies are both modeled and managed within the controller rather than as separate Kubernetes objects

• In Kagenti agent dependencies are modeled via an additional resource and operator labeled Platform where another set of Component CRDs with its own operator may be explicitly specified. Dependencies are modeled and reconciled as Kubernetes objects.

At a glance, this would appear to make Kagenti more flexible while Kagent is more monolithic, but with less work required in order to define common relationships since the state is managed via a single controller.

Here we can discuss which project might better align with our goal of offering an "Agent" as first class resource in openshift.

[usize]

Here's an example of a missing feature that integration with openshift would solve (via routes) kagent-dev/kagent#536

[zanetworker]

One issue here I could see is that Kagenti makes assumptions on how many platform primitives should be built, there seems to be little means to delegate functions to "3rd party" for things like auth, governance,... on the other hand its a good turn-key solution. While kagent focuses on just solving the Agent Lifecycle problem (also fewer CR dependencies/yaml).

In both cases I believe we need to make sure we tackle the right personas and builds stories for each:

• AI Users -> Consumers for already built experineces/agents/AI systems (built by the AI Engineers)
• AI Engineer (wants abstractions cares less about CR/CRDs, aka consumers of APIs curated/exposed by the plaftormers/Opsers)
• Platform/Ops -> Main persona for this here eposing a construct for packaging AI Applications

[usize]

Very good point, Kagenti's "enterprise ready" approach, where it integrates tekton, spire, keycloak and istio has some advantages for a platform/ops persona but has a number of downsides. The overall trend is, it solves problems comprehensively, but at the expense of simplicity.

I see some advantages in terms of being able to quickly stand up a complicated architecture since a lot of difficult problems are pre-solved, but they also put pressure on the components you bundle to maintain parity with competing projects and increase the minimum-complexity of deployments quite substantially. Is it worth it?

With kagent, I got everything stood up very quickly and easily but it's a barebones example. I enjoy the smaller set of assumptions and the ease.

I need to spend a little more time to compare the projects and think about who's going benefit most from one approach over the other. I'll write up what I find.