The first time a network blip cost me a Hermes session, I was twenty minutes into a refactor and the agent had built up a map of the codebase in its head that I did not have. The gateway lost its connection. The session ended. The map went with it.
I am not going to claim that exact experience is the reason v0.13.0 happened. The release notes call this one "the tenacity release," which sounds like a phrase the team picked out of a thesaurus until you realize three of the headline features in v0.13.0 are different answers to the same question: how does an agent finish what it started?
May 7, 2026, exactly one week after v0.12.0. The release dropped on a Wednesday. By the end of that Wednesday, the agent on my server had three new ways to be stubborn.
A Kanban board, but for AI workers
The marquee feature is the Multi-Agent Kanban — a durable task board where multiple Hermes workers can pick up, hand off, and complete tasks collaboratively. The infrastructure under it is not pretty, but it is the part that matters: heartbeat monitoring, zombie detection, per-task retry budgets, atomic state writes.
The plain-English version: if a worker dies mid-task, another worker can claim it. If the gateway restarts, the board comes back where it was. If a task fails three times in a row, the retry budget says enough, log it for a human. The point of all this plumbing is that you can point five copies of Hermes at a backlog and walk away.
The other half of the feature, less flashy but probably more useful for a single-instance deployment: a single Hermes instance can itself behave as a "kanban worker" in a multi-process plan, picking up tasks one at a time. The agent has, in some sense, learned to be a queue consumer. That changes what kinds of jobs you can reasonably hand it.
/goal and the Ralph loop
The /goal slash command is the implementation of what the agent-design community has been calling the Ralph loop — a long-running design pattern of telling an agent to stay locked on an objective across multiple conversation turns, no matter how the user steers in any single turn.
If you have ever caught Hermes drifting from the original task because the most recent user message pulled it sideways — "wait, what were we doing again?" — /goal is the fix. You set the goal once at the start of a session, and every subsequent turn is evaluated against that goal. The agent is allowed to take detours; it is not allowed to forget the destination.
/subgoal, which lands a week later in v0.14.0, lets you bolt on success criteria mid-run. The pair together is the closest a chat-shaped agent has come to what a traditional project plan does: a target, with refinements over time.
Session durability — the gateway restart is no longer a reset
The third "finishes what it starts" feature is session durability. Conversations auto-resume after gateway restarts, preserving context and thread routing. Under the hood, this is checkpoints v2 plus atomic session persistence — the same shape of work that powers the Kanban's heartbeat layer.
The user-visible behavior is the bullet I would have killed for several months ago: a gateway restart is no longer a conversation reset. Pull a new release, reboot the process, the conversation picks up where it left off. The day-2 operational story for self-hosted Hermes just got a lot simpler — and for anyone running it on a VPS that occasionally reboots itself for kernel updates, this is the release where the agent stops minding.
Google Chat, DeepSeek v4 Pro, and the model wave
Google Chat lands as the 20th messaging platform. The release also introduces a pluggable provider system so third-party adapters can ship outside the core repo — the same shape as the v0.11.0 transport architecture work, now extended to gateway adapters. The next iMessage clone, the next regional messaging API, the next encrypted chat app — none of them has to wait for a core merge anymore.
On the model front: DeepSeek v4 Pro, xAI Grok 4.3, OpenRouter Owl-Alpha (free tier), and Tencent Hy3 Preview are all new entries. xAI Custom Voices ships with voice cloning support for TTS. And a video analysis tool rides on top of Gemini and other multimodal-capable models, finally giving Hermes a way to look at a 30-second clip instead of a stack of stills.
If your benchmark for "is this project keeping up with the model frontier" is "did they ship support for the model I tried this week," v0.13.0 almost certainly passes.
Internationalization arrives
The CLI and documentation got translated into seven locales — Chinese, Japanese, German, Spanish, French, Ukrainian, Turkish. That is a one-line bullet that buries a lot of work. If you have been reading Hermes's CLI output in English for the last three months, and your brain treats English as a second language while you are tired, v0.13.0 is the release that lets you switch.
A side effect worth pointing out: when CLI output is localized, error messages get localized too. Which means a non-English-speaking user can paste an error into a forum and the responders see the same words. Translation is, surprisingly often, also a debuggability story.
Security, ACP, and the curator's archive subcommand
Eight P0 vulnerabilities closed — the release notes do not enumerate them by CVE, but the pattern is broad: redaction enabled by default again (reversing v0.12's flip, now that the redactor itself has been reworked), Discord role allowlists scoped to guild, WhatsApp rejecting unknown contacts by default, TOCTOU windows closed in auth systems.
hermes curator archive, prune, and list-archived formalize the v0.12 curator into proper subcommands. /steer and /queue show up in the ACP adapters (VS Code, Zed, JetBrains) — if you are using Hermes from an editor, you can nudge a running agent from inside the editor now, not just from the terminal.
What "tenacity" actually buys you
Each of these features individually has been in the release tracker for a while. Together they buy you the same thing: a Hermes that does not forget what it was doing when the network blinks, the model API throttles, the gateway restarts, or the user gets distracted.
The agent has always been smart. v0.13 made it stubborn. There is a difference between the two. Smart agents recover from interruption gracefully — they often start a new task with no memory of the old one. Stubborn agents finish the old task. The Kanban, the goal lock, the session resume — they are all the same shape of feature: state that survives the things that used to destroy it.
---
Most of the AI-agent industry, this quarter, is competing on raw capability — bigger models, longer contexts, more skills. v0.13.0 is competing on a different axis. The hard problem for self-hosted agents in 2026 is not "can it do X." It is "can it do X for thirty consecutive minutes through a network event." A surprisingly large fraction of agent-quality-of-life turns out to live in the answer to that second question.
The fact that the release is named "tenacity," rather than something with the word durability or reliability in it, is telling. Tenacity is a word people use about other people. The implication is that the team thinks of the agent the same way.