I picked the recent Hermes MCP catalog safety research thread because it had the exact shape Kanban is good at: parallel investigation, synthesis, review, publishable output, and a need to keep uncertainty visible.
The work had multiple lanes: discover the landscape, evaluate risks, map findings back to Hermes, synthesize a recommendation, and possibly publish or preserve the result. Those lanes can run in parallel but should not all collapse into one opaque final answer.
This page uses the recent thread at a process level, not as a dump of private conversation contents. The point is to show how a similar research workflow could be structured on a Kanban board.
The parent agent chooses the research lanes and starts workers in parallel.
If a lane is slow, fails, or needs clarification, the state is mostly buried until the result summary returns.
The final answer may be good, but the intermediate decisions, source quality, and uncertainty are not durable first-class artifacts.
Verification, live URL checks, and follow-up pages happen in the same chat thread rather than as explicit dependent tasks.
Create separate research cards for ecosystem inventory, threat model, policy controls, and Hermes implementation mapping.
Create a synthesis card that depends on those parent tasks. It starts only when the research handoffs exist.
Create a review/publish card that depends on synthesis and requires explicit verification before a public page or operational change ships.
You could add a comment mid-run — “also look at permission prompts” or “do not publish internal details” — and the relevant worker or synthesis card would carry that context.
Each research card could complete with sources read, confidence level, unresolved questions, and “what I did not check.” The synthesis card would not have to reverse-engineer that from prose.
Instead of “Donn is working,” the board could show “threat model done, Hermes mapping running, policy choice blocked.” That is a better map of reality.
A publish card could require checks: no secrets, no private operational context, HTML parses, GitHub Pages returns 200, expected text marker present.
Once the card template works, the same research pattern could be reused for tools, plugins, providers, security controls, or buying decisions.
If a worker times out or hits a credential/tool issue, the task can be blocked or retried without losing the rest of the workflow.