Jonathon's AI Wiki

Hermes Multi-Agent Workflow — Kanban-Coordinated Triage Pipeline Template (tonbistudio)

5 min read

Source: ai-research/hermes-multi-agent-workflow-tonbistudio-readme-2026-06-03.md (github.com/tonbistudio/hermes-multi-agent-workflow README), raw/gh-star-tonbistudio-hermes-multi-agent-workflow.md

tonbistudio/hermes-multi-agent-workflow is a reusable, open-source (MIT) skeleton for an autonomous, human-gated multi-agent triage pipeline built on Hermes Agent. A fleet of agents detects items from sources, dedups, scores them against a rubric, researches them in parallel, routes each to a fulfillment path, pauses at one human approval gate, then fulfills and delivers — all coordinated on a single Hermes Kanban board. It ships pre-wired as a worked example (find AI-agent pain points → build a fix or explainer video) you can read end-to-end and then repoint at your own domain by editing one config file. It is a template, not a turnkey app — it passes its tests and validates config out of the box, but going live needs your own Hermes install, profiles, auth, and scouts.

Key Takeaways

  • Fixed pipeline shape, swappable contents. sources → intake → dedup → score → research (parallel) → route → [path A / path B / shelve] → HUMAN GATE → fulfill → deliver. The shape is fixed; everything domain-specific lives in one triage.yaml (sources, rubric, research lanes, route map, paths, roles).
  • Kanban board as the message bus. A single Hermes Kanban board does triple duty — bus, dispatcher, and fan-in. Agents read/claim/move cards rather than calling each other directly (docs/02-the-board.md). This is the concrete orchestration substrate behind the “Hermes as orchestrator” idea.
  • Hard config/code split (“fat engine / thin skill”). Generic engine code in engine/ (config.py, engine.py, scoring.py, routing.py, dedup.py, item_vault.py, kanban_store.py, intake_parser.py, proposal_actions.py) is rarely edited; new domains go in triage.yaml + markdown templates, never the code. The contributing rule enforces keeping engine/ domain-agnostic.
  • One explicit HUMAN GATE. A single approve / shelve / modify checkpoint sits between routing and fulfillment (proposal_actions.py, config-driven) — a concrete implementation of the human-in-the-loop approval other agent articles describe abstractly.
  • Scoring + dedup are real engine stages. Rubric scoring runs in LLM mode or deterministic mode; dedup uses token-cosine similarity (embedding-ready); a per-item cost report (scripts/cost_report.py) feeds a cost gate. ^[the README states these modules exist; their internal algorithms were not read at ingest]
  • Adapt-via-coding-agent built in. Hand your coding agent the repo’s AGENTS.md and have it walk docs/04-adapting-to-your-domain.md: edit triage.yaml → edit paths/ templates → edit skills/templates/ (scout + orchestrator SKILL.md) → python -m cli.triage validate → keep tests/ green → docs/07-runbook.md to go live.
  • Genuinely runnable scaffolding. pip install -r requirements.txt (just PyYAML), python -m cli.triage validate, python -m unittest discover -s tests (12 generic tests), python -m cli.triage scaffold (prints the Hermes setup plan). CLI verbs: validate / scaffold / init / install.
  • Verified repo, not a thin demo. MIT license, 7 deep-dive docs (docs/01–07), real tests, a worked reference example (examples/ai-agent-pain-points/REFERENCE.md), SECURITY.md + docs/06-security.md (it runs LLM-authored code and shells out behind the gate), and CONTRIBUTING.md. 153 stars, Python, pushed 2026-06-01; maintainer tonbistudio is an established AI/ML researcher (separate ~1K-star repo, build-in-public channel).
  • Second worked example — LLM-wiki maintenance (2026-06-09 video). [YouTube signal — tonbistudio 2026-06-09] The maintainer published a hands-on follow-up applying this same template to maintaining/updating a Karpathy-style LLM wiki (knowledge base) — pitched as “more generally applicable” than the bundled AI-agent-pain-points example. It’s a live build-from-template demo (what to change in triage.yaml + the design of the wiki-maintenance pipeline). Directly meta-relevant: a Hermes multi-agent Kanban workflow for maintaining exactly the kind of wiki this project is. (Source: raw/How_to_Build_a_Multi-Agent_Workflow_for_LLM_Wikis_in_Hermes_Kanban.md, https://www.youtube.com/watch?v=hbKvO5MWq08.)

Why It Matters

  • The missing “how do Hermes agents coordinate” answer. The wiki documents Hermes orchestration conceptually (complexity-routed SWE workflow also uses a Kanban board), but this is a readable, testable reference implementation of the Kanban-as-bus pattern with fan-out / fan-in / human-gate structure.
  • Config-over-code is the reusability lever. Pushing all domain logic into triage.yaml + templates (engine stays generic) is the same discipline that makes skill bundles and SKILL.md templates portable — and it’s what lets a coding agent adapt the whole pipeline by editing one file.
  • The HUMAN GATE is a transferable safety primitive. A single, explicit approve/shelve/modify stage between automated routing and action is a clean pattern for any autonomous pipeline, and it pairs with the Hermes security model (it runs LLM-authored code + shells out, so the gate + scope rails matter).

Implementation

Repo: github.com/tonbistudio/hermes-multi-agent-workflow | License: MIT | Language: Python | Stars: 153 (2026-06-03) | Built on: Nous Research Hermes Agent.

Setup (read-first):

  1. pip install -r requirements.txt; python -m cli.triage validate; python -m unittest discover -s tests.
  2. Read examples/ai-agent-pain-points/REFERENCE.md to see the complete worked pipeline.
  3. To adapt: edit triage.yaml + paths/ + skills/templates/; re-validate; follow docs/07-runbook.md for profiles / board / crons / go-live.

Cost: Free (MIT). Running it consumes Hermes model/tool usage (e.g. via Nous Portal) at the cost of whatever models the scoring/research stages call.

Integration notes:

  • Read SECURITY.md + docs/06-security.md before deploying — it runs LLM-authored code and shells out behind one gate; there’s a pre-publish secret-scan checklist before open-sourcing an adapted copy.
  • Same maintainer (Tombi Studio / “On Chain AI Garage”) publishes the Hermes-driven Hyperframes video walkthrough — the “make an explainer video” fulfillment path connects the two.

Try It

  1. Clone, pip install -r requirements.txt, run validate + the 12 tests + scaffold to read the setup plan without committing to a live Hermes install.
  2. Read examples/ai-agent-pain-points/REFERENCE.md and docs/02-the-board.md to learn the Kanban-as-bus pattern, then sketch your own triage.yaml (sources + rubric + route map) for a domain you actually triage.

Open Questions

  • The internal algorithms (scoring rubric math, routing classification, token-cosine dedup thresholds) were not read at ingest — only the README’s module list. Verify against engine/ + docs/03 / docs/05 before quoting specifics.
  • How the parallel research stage fans out across Hermes sub-agents / worktrees (vs the Kanban Swarm in the v0.15 release) — not detailed in the README.
  • Adoption is early (153★, single maintainer) — durability / maintenance cadence unproven.

Graph View

Backlinks