The time is finally ripe for multi-agent systems, e.g. agent-to-agent (A2A). This has been the hype for the past year, but finally the pieces are coming together to make it useful. I don't mean the simplistic use of subagents in Claude or simply having different roles using prompts.
I mean creating multiple autonomous / independent agents with their own identity, memory, permissions and workspaces. To make this work in a secure way you first of all need a proper "governance" system, e.g. agent sandboxes which allows different agents having different policies and their own computers. Our own product here is Lens Agents by Mirantis.
Then, you need a "harness" that is designed to be autonomous, which includes a memory and identity system, not a simple "you are role x" prompt. Think of basically what OpenClaw was doing, but nowadays there are better options from most players in the space (at Lens we have our own Prism agent, of course).
The actual protocol of doing the communication between agents doesn't really matter. It can even be done through Slack - or through a custom implementation. These enterprise "open" standards like A2A are not really important here (and who even uses those?). Implementations are cheap which makes protocols less useful.
The whole multi-agent platform should also be running remotely, not on developer machines. It should also be "multiplayer", e.g. multiple humans should be able to interact with these agents. The point is not creating an army of agents running on your laptop.
If you have all that up and running you are ready to create a state of the art software factory.
Top comments (0)