Sketch notes on multi-agent pipelines vs. human task queues
Quick answer: Kanban inside Hermes is a multi-agent work queue. Not your todo list. It's where autonomous workers claim tasks, run research, and build deliverables while you sleep. Todoist is for your actions. Kanban is for Tina's pipeline.
What It Is
Hermes has a built-in Kanban board (SQLite-backed) that lives alongside sessions and cron jobs. Tasks are not reminders — they are units of work assigned to agent profiles:
orchestrator
→
researcher
→
analyst
→
writer
→
deploy
Each worker is a separate Hermes profile with its own context and toolset. The orchestrator breaks work into cards. Workers claim them, execute, and mark done.
Live Board (Right Now)
Ready
Synthesize: Agentic SaaS research findings
t_fcccbeae · analyst
Todo
Draft: Agentic SaaS research brief
t_82ae71da · writer
Blocked
Overnight Planning — 6PM Orchestrator
t_26d47fa3
Research competitor pricing changes
t_b616a578
Done
Research: ServiceNow AI agents
researcher
Research: Atera agentic features
researcher
Research: SuperOps.ai agentic
researcher
+ 12 more ITSM tasks...
This is real data from the board as of today. The ITSM/MSP weekly research pipeline feeds into this queue every week.
Kanban vs. Todoist vs. Cron vs. Manual
System
What it holds
Who acts
Trigger
Kanban
Agent work units (research, synthesis, draft, deploy)
Agent profiles (orchestrator, researcher, writer)
Cron or manual kickoff
Todoist
Human action items, deadlines, approvals
You (Prabha)
You add tasks manually
Cron
Scheduled triggers (6PM scan, 3AM run)
System clock
Time-based
Manual chat
One-off requests, quick questions
You + Agent Tina
Direct message
Kanban does not replace Todoist. They are layers. Todoist = operational truth for human work. Kanban = execution truth for agent pipelines.
My 4 Boards
Tina Board — Auto-run everything. Optimization, error logs, setup fixes, model memory issues. No approval gate. If it's broken, Tina fixes it.
Professional — Auto-research is allowed. Auto-build design/code needs approval. Research flows into HTML artifacts or Google Docs. Design tasks wait for your thumbs-up.
App Work — Auto-plan and research. Builds and App Store releases need explicit approval. WhatPay screenshots, ASC pushes, metadata updates live here.
Personal — Manual/dormant. No auto-execution. You add what you want, when you want.
How to Use It
hermes kanban list — Show all tasks and states
hermes kanban show t_fcccbeae — Inspect one task
hermes kanban create "Research: New ITSM player" --role researcher — Add a task
hermes kanban complete t_fcccbeae — Mark done
Tasks in Kanban are not cron jobs. They are durable work units that survive across sessions. A cron can kick off a Kanban task; a Kanban task can spawn a cron. But they are separate systems.
Why This Exists
Before Kanban, everything was either:
Cron-only — rigid time slots, hard to chain dependent work
Delegate-only — subagents die when parent session ends
Manual-only — you had to babysit every step
Kanban adds a stateful middle layer: work is decomposed into cards, cards wait in queues, agents claim and execute them independently, and the board tracks what's ready, what's stuck, and what's shipped.
For ITSM/MSP weekly research, the orchestrator creates research cards → researchers run in parallel → analyst synthesizes → writer drafts the HTML brief → deploy pushes to thisisprabha.com/sketch/. The whole pipeline runs without waking you up.
Current Status
The Kanban system is live and working. The ITSM research pipeline has completed 15+ tasks across ServiceNow, Atera, HaloPSA, Freshworks, SuperOps.ai, and emerging players. Two tasks are blocked (likely due to the old overnight orchestrator still being referenced). One synthesis task is ready to claim. One draft task is waiting behind it.
If you want the weekly ITSM research to actually produce the HTML brief, claim the ready synthesis task or unblock the orchestrator.