Miadi Review

[TEAMS][COORD]Hermes Multi-Agent Kanban Pipeline Engineering Review

Sep 22, 2026 · model: manual-edit · channel: Tonbi's AI Garage

Watch source

Hermes Multi-Agent Kanban Pipeline Engineering Review

Source: I Built the Ultimate Multi-Agent Workflow w/ Hermes Agent Kanban Board — Tonbi's AI Garage https://youtu.be/EKVRqcpTT6s

Intent

William

  • Have an adequate ways to create teams of agents that works in their adequate level as part of the "Relational Software Factory" and serves us well
  • Reservations would be that we can't use a claude-code subscriptions but that does not mean that we can't have workers or coordinators

Summary

Tonbi demonstrates an autonomous multi-agent pipeline orchestrated on top of Hermes Agent using a durable SQLite-backed Kanban board. The system uses web and social scouts to detect developer pain points, runs automated rubrics and parallel research agents, routes proposals through a Telegram-based human approval gate, and autonomously implements runnable software tools or educational slide decks.

related reviews:

review/09de3362-2418-4172-8250-fdcd913c0aa0

https://miadi-review-service.vercel.app/review/09de3362-2418-4172-8250-fdcd913c0aa0

  • This would provide an Approach to create the teams

Related explorations

miadi-swarmclaw

  • miadi-swarmclaw and all of its potential services are being explored and one of them is a form of community where the agents are capable to interact which is nice....

Aspects

Durable SQLite Board as Coordination Bus (04:04-06:26)

Instead of volatile message queues or inter-agent chat, tasks exist as durable cards stored in a single SQLite database. The dispatcher loop atomically claims ready tasks, spawns agent profiles in isolated workspaces, and preserves full state across crashes and system restarts.

Event-Driven Fan-Out and Fan-In Choreography (05:36-06:26)

Dependent tasks automatically stay queued until upstream parent tasks finish. When an issue enters research, three researcher agents execute in parallel to verify sources, extract context, and audit solutions, automatically firing the downstream routing stage once all three complete without polling code.

Autonomous Pain-Point Scraping and Viability Rubric (07:22-09:30)

Scout agents continuously monitor platforms like X, Reddit, and GitHub for developer issues. The orchestrator deduplicates submissions and evaluates them across a 100-point rubric assessing frequency, intensity, solvability, and strategic fit, automatically shelving any task scoring below 65.

Single-Reply Telegram Human Gate (10:16-11:08)

Human intervention is restricted to a single approval step before build execution. The orchestrator delivers synthesized proposals directly to a Telegram channel, where one-tap approve, shelve, or modify commands trigger downstream execution while keeping token expenditure strictly bounded.

Self-Healing State and Asset Recovery (25:48-27:14)

When downstream agents encounter transient workspace faults or lost scratch artifacts during build steps, self-healing recovery tasks automatically trigger. The pipeline regenerates missing deliverable files into persistent directories and resumes execution without human debugging.

End-results and Core Aspect

Core Aspect — End-to-End Autonomous Problem Remediation (11:08-12:06)

The architecture takes real-world developer friction from initial web discovery to verified deliverable generation, requiring human attention solely to approve evaluated proposals via chat.

Results

  • Automated Config Verifier CLI Generation (28:26-29:47): In response to VS Code extension configuration issues, the builder agent generated a standalone Python CLI tool, automated unit tests, and structured markdown diagnostic playbooks within persistent storage.
  • Educational Video Deck and Script Production (30:33-32:00): For complex subagent MCP tooling failures, the producer agent compiled an 18-slide standalone HTML presentation alongside chapter outlines, narration scripts, and technical fact sheets.