Hermes Agent v0.13.0 "Tenacity": Multi-Agent Kanban, /goal Command, and 12 Key Features

Key Takeaways

• Multi-agent Kanban board: tasks drop into columns, AI workers pick them up with heartbeat tracking; a hallucination gate catches workers who claim to finish tasks they actually didn't complete.
• /goal command locks the agent onto a persistent objective across turns, solving the common problem of agents drifting from the original task.
• Session auto-resume: conversations survive gateway bounces, agent update restarts, and source file reloads — no lost progress.
• Checkpoints v2 rewrites state storage with real pruning and disk guardrails — "no more orphan shadow repos."
• Post-write delta lint: Python, JSON, YAML, and TOML files are linted immediately after every write operation, surfacing syntax errors before they propagate.
• Eight security issues closed: D-Action on by default, Discord role allow-lists now guild-scoped, WhatsApp rejects stranger messages by default.
• New model support: DeepSeek V4 Pro, Grok 4.3, Owl Alpha, and Tencent HY Preview all added via OpenRouter.
• Google Chat becomes the 20th supported messaging platform; voice cloning via ElevenLabs custom voices added.
• No-agent cron mode for background monitors: script-only cron jobs that only ping the agent when there is something to handle.
• Seven new language locales: Chinese, Japanese, German, Spanish, French, Ukrainian, and Turkish.

Multi-Agent Kanban Board

Think of it like a Trello board where the cards are moved by AI workers instead of people. You create a board, drop tasks into columns (triage, todo, ready, in progress, blocked, done), and multiple Hermes worker instances pick them up, execute them, hand them off, and complete them. Workers use heartbeats so the system knows they are still alive — if a worker goes silent or exits without completing, the system reclaims the task and can give it to another worker.

The hallucination gate is the critical addition. One of the persistent failure modes in AI agents is a worker that reports finishing a task without actually finishing it — it claims cards exist that don't. The hallucination gate detects this and triggers recovery. You can run multiple boards from one install, share boards across profiles, set per-task retry budgets, detect zombie workers, and auto-block workers that exit without completing. This is a real project management system built for agents rather than humans.

The /goal Command

When you give an AI agent a complex, multi-turn task, it tends to lose the original objective after a few steps — the context window drifts, the agent starts optimizing for the current sub-task rather than the stated goal. /goal is a first-class primitive that locks a persistent objective into the agent's working context. A dedicated docs page now covers the pattern in detail. Combined with the Kanban system and session durability, this release addresses three of the most common reliability failures in AI agent work: goal drift, false completion reports, and state loss on restart.

Commands & Code Mentioned

/goal
/model
hermes dashboard

Related on OpenClawDatabase

• Hermes Agent Setup Guide
• Hermes Tasks — scheduling and managing automated agent tasks
• Hermes Memory — how Hermes stores skills and project context
• Hermes vs OpenClaw — platform comparison
• Hermes MCP Tools