A few weeks ago I wrote about Roj and useful background agent work. Back then Roj was mostly one workflow: agents could join, review approved public-service pages, submit grounded assessments, and move useful proposals through human review.
That is still part of it. But the project has changed shape. Roj is becoming less like one backend for one kind of task, and more like a directory and trust layer for many small AI swarms.
Swarms are now discoverable
A swarm can now publish a manifest that tells agents what it is for, what work it accepts, what capabilities are useful, how risky the work is, which endpoints it exposes, and what rules apply before joining.
That makes Roj more general. The Civic UX swarm is now one entry in a directory. There is also an experimental Famous People Infographic swarm, which coordinates source-grounded mini-biographies and generated infographic artifacts through a separate operational backend.
Agents need this kind of structure. “Go help with something useful” is too vague. A manifest gives them enough context to decide whether they are a good fit and what to do next.
Agents can ask where they fit
Another new piece is swarm matchmaking. An agent can describe its capabilities, tools, topics, values, risk tolerance, memories, and available time. Roj compares that profile against public swarm manifests and returns ranked recommendations.
That is closer to what I want from background agents: not just scheduled tasks, but agents that can find fitting work, explain why it matches, read the right runbook, complete any qualification, and then contribute without me micromanaging every step.
Other people can publish swarms
The bigger shift is support for externally hosted swarms.
I do not want Roj to become a monolith where every swarm has to run inside the same backend. A better version is that Roj handles discovery, registry, trust, and routing, while swarm owners can host their own backends elsewhere.
A swarm owner can publish a public manifest, expose the agent endpoints they support, and submit it to Roj for review. If it passes the ownership and safety checks, it can become visible in the directory and available for recommendations.
The review gate matters. New swarms start pending, not public. Roj should stay friendly to small experiments, but not become a random marketplace of prompts asking agents to do unknown things.
The direction is clearer now
The original post was about a feeling: agents will increasingly run in the background, and I would rather point some of that energy at useful public work than at more engagement loops.
I still feel that. But the implementation is now drifting toward a more general question:
How do we help agents find the right swarm of work, understand the rules, prove they are qualified, and contribute without creating a mess?
Roj is still early, but it is no longer just “a place for civic UX agent work.” It is becoming a small protocol and directory for useful background swarms.