Repository navigation
Show & tell: running A2A agents on Kubernetes with Kynomesh #2350
whynowy
started this conversation in
Show and tell
Replies: 2 comments 1 reply
|
BTW - I will speak at KubeCon Salt Lake City, Co-located Events, [Agentics Day: MCP + Agents], to talk about Orchestrating, Discovering, and Autoscaling Distributed Multi-Agent AI Systems with Kynomesh, hope we can meet there. |
0 replies
|
Drift reload is a great idea. One gap it exposes: when a peer's card changes, the supervisor rebuilds its tools from the new card and trusts it immediately. A hook that verifies the new card's skills behave as declared before the reload completes (or at least flags the change) would stop a bad deploy from propagating through the whole AgentSet. Is there a place in the reload path for that kind of check? Kasper |
1 reply
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
Hi everyone,
I'd like to share Kynomesh, an open source project that runs A2A agents on Kubernetes, and ask this community for feedback.
Repo: https://github.com/kynoproj/kynomesh
Doc: https://kyno.sh
Blog: https://blog.kyno.sh
What it is
Kynomesh lets you declare a group of cooperating agents as one Kubernetes resource, an
AgentSet, and takes care of running them and wiring them together over A2A. Each agent runs in its own pods with an injected broker sidecar that handles A2A transport. Your agent code stays an ordinary A2A agent.A few things it does:
Supervisor,HandofforSequential) and an entry agent, and Kynomesh works out who can call whom. Agents call peers by name, with no URLs in code.AgentCards. When a peer's card changes, Kynomesh can restart only the agents still holding the stale copy (we call this drift reload).AgentSetthat mixes both.AgentSetcan reference an agent that lives in anotherAgentSet, or at any URL, and call it like any other peer.Try it
Where it is today
It's early and under active development, and minor versions can contain breaking changes. We'd like to hear frosystems before we settle more of the design.
Feedback we'd value
Supervisor,HandoffandSequentialcover how you connect agents? What topologies are yo "handoff" mean precisely?Issues, discussions and PRs - I'm happy to answer questions here.
Thanks!
All reactions