I think agents should work like real teams, with independent, distinct roles, async communication, and the ability to onboard new teammates or tools without restructuring the whole org. I built backend systems at Yahoo and TikTok so event-driven agents felt obvious. But no agent SDKs were using this pattern, so I made Calfkit.
Calfkit breaks down agents into independent services (LLM inference, tools, and routing) that communicate asynchronously through Kafka. Agents, tool services, and downstream consumers can be deployed, added-to, removed, and scaled independently.
Check it out if this interests you! I’m curious to see what y’all think.
I think there's a big asymmetry between the kind of user and their environment that wants to toy with agent teams and the kind of user that would ever want to deal with the unwelcome hassle of having kafka as a dependency.
Possibly. However I’m not particularly tied to Kafka either so I wouldn’t be opposed to adding compatibility with alternative lighter weight options later on.
Has anyone actually built event-driven agents yet? Or any agents, outside some basic dev workflows? (PR review, etc)
I’d argue that many new AI startups are simply agents built on top of existing, established workflows. With today’s agent SDKs and AI coding tools, building them is incredibly easy.
But as people ship faster, often without understanding scalable system design, we’re heading toward an era of slow, fragile, and unscalable Saas. I believe that eventually products built on solid infra early will outlast the wave of slop--like how facebook outlived friendster. That's why I built Calfkit.