Hermes Kanban Enables Durable Multi-Agent Workflows
Hermes v0.11/0.12 shift from chat agents to persistent systems via Kanban boards: local SQLite tasks with dependencies, structured handoffs, retries, blockers, and crash recovery for workflows like feature shipping or PM-engineer-reviewer pipelines.
Persistent Coordination Over Ephemeral Delegation
Hermes distinguishes short-lived delegation (function-call style sub-agents that return immediately) from Kanban work queues for durable, multi-role workflows. Kanban tasks persist in a local SQLite database (hermes/con.db), shared across profiles, with fields for status (Triage, Todo, Ready, In Progress, Blocked, Done), assignee, parent/child dependencies, comments, run history, and structured handoff data. Dependencies auto-promote child tasks upon parent completion, preventing premature execution—e.g., API implementation waits for schema design, tests wait for API. Handoffs carry summaries and metadata (e.g., changed files, decisions) to downstream agents, avoiding chat log digging. Use delegation for quick subtasks; Kanban for cross-boundary work needing restarts, human input, or audits.
v0.11's pluggable transport layers enabled broader providers (AWS Bedrock, NVIDIA NIM, Grok API, Google Gemini, Versel AI Gateway, GPT-4.5 via Codex) and smarter delegation with orchestrator sub-agents. v0.12's autonomous Curator grades/prunes skill libraries on schedule; upgraded self-improvement loops use rubric-based reviews, prefer updating recent skills, handle references/templates, inherit parent runtime. Providers expanded (GMI Cloud, Azure AI Foundry, Mistral O1, Tencent TokenHub, LM Studio); gateways added (Microsoft Teams, WeCom); tools bundled (Spotify, Google Meet, ComfyUI, TouchDesigner). Dashboard gains models tab; 57% faster 2e cold starts; local Piper TTS.
Four Workflow Patterns for Shipping Work
Solo feature shipping: Chain dependent tasks (design schema → implement API → write tests). Completion handoffs metadata like DB tables or files, ensuring context flows without re-researching.
Fleet farming: Queue independent tasks for specialist profiles (translator, transcriber, copywriter). Dispatcher assigns via embedded gateway; lanes-by-profile view tracks parallel progress, with handoffs for analytics (e.g., tokens translated).
RDO pipeline with retries: PM specs → engineer implements → reviewer checks. Blocks on feedback (e.g., missing password check); unblock/retry preserves run history (outcomes, summaries, metadata per attempt). Reviewers access parent summaries/files before diffs, mimicking real engineering.
Dispatcher commands: hermes kanban launches dashboard with filters (search, tenant, assignee), lanes toggle, nudge button for immediate dispatch ticks.
Crash Recovery and Scoped Reliability
Circuit breakers limit retries on spawn failures (e.g., missing API keys), marking tasks Blocked with 'gave up' to avoid infinite loops. Mid-task crashes (OOM, network) release claims, revert to Ready for fresh workers; history logs issues (e.g., 'crashed: OOM' → 'completed: chunked strategy'). Single-host design (local SQLite, same-machine workers) suits personal coordination, not multi-server enterprise—expose dashboard cautiously (avoid 0.0.0.0). v0.11's Ink-based TUI adds sticky composer, live streaming, status bar, light theme; SL steer nudges post-tool-call; extensible dashboard/plugins.
This builds production-grade agent systems: visibility into stuck tasks, failure traces as data, role handoffs with context—far beyond chat logs.