Hermes ships desktop app to run persistent Kanban agents locally

AI agents keep running into the same wall. They look impressive in a chat window, but the minute you ask them to juggle real work over hours or days, the whole thing gets fragile. Context drifts, subprocesses die, and “multi-agent” often means a bunch of temporary helpers that vanish when the parent call ends. Hermes Agent is trying to fix that by turning the agent into a durable system instead of a clever prompt wrapper. Over the last two releases, and especially in the May 7 v0.13 update, Nous Research pushed Hermes much harder in that direction. ### What actually shipped? The big change is Kanban. Hermes now has a durable multi-agent board where tasks can be created, claimed, handed off, blocked, retried, and completed over time, instead of being trapped inside one parent-child agent call. The docs describe it as a shared task board across Hermes profiles, and the v0.13 release frames it even more bluntly: “delegate to an AI team that actually finishes.” ### Why is that different from normal “subagents”? (github.com) Because these workers are not just little disposable branches. Hermes says each worker is a full OS process with its own identity, and every task transition is written to a SQLite database at `~/.hermes/kanban.db`. That is the important architectural shift. It means the workflow behaves more like a queue plus state machine than a one-shot RPC call. If one worker dies, the whole plan does not disappear with it. (hermes-agent.nousresearch.com) ### What did v0.13 add on top? Durability features. v0.13 added heartbeats, reclaim logic, zombie detection, retry budgets, auto-blocking on incomplete exits, and a “hallucination gate.” It also added gateway auto-resume, so interrupted sessions can come back after a restart, plus `/goal`, which keeps an agent locked onto a target across turns. Basically, Hermes is patching the exact failure modes that make long-running agents annoying in practice. (hermes-agent.nousresearch.com) ### So where does the desktop angle fit? The core Hermes project itself is still centered on a local gateway, CLI, dashboard, and docs-driven install flow, not a single official polished desktop binary from Nous. But the product surface is clearly moving toward “local app” behavior — install it on your machine or server, manage it through a dashboard, and let it keep running in the background. There are also third-party desktop wrappers already forming around the Hermes runtime, which tells you the ecosystem sees the same opening. (github.com) ### Why does model routing matter here? Because Hermes is not trying to be married to one foundation model. The README pitches provider switching with no code changes and lists a long menu of backends, from OpenAI and Anthropic-style APIs to OpenRouter and custom endpoints. That matters more in a Kanban setup than in a chatbot, since different tasks can be pushed toward different models, prices, or capabilities without rebuilding the workflow around one vendor. (hermes-agent.nousresearch.com) ### Is this just for coding? No — and that is part of the pitch. The Kanban docs call out research triage, scheduled ops, digital twins, engineering pipelines, and fleet work like managing many accounts or monitored services. In other words, Hermes is positioning itself less like “AI pair programmer” and more like a persistent operations layer that happens to speak natural language. ### Why are people paying attention now? (github.com) Scale helps. The repo is sitting around 147,000 GitHub stars, and the release cadence is fast enough that the project feels alive rather than aspirational. More important, the new features are not cosmetic. They are the boring, load-bearing pieces — persistence, recovery, retries, security defaults, platform support — that make unattended agents feel less like demos and more like infrastructure. ### What is the real takeaway? (hermes-agent.nousresearch.com) Hermes is making a bet that the winning AI interface is not one chat box. It is a local, persistent workflow system where agents keep state, coordinate on boards, survive crashes, and keep going while you are away. That is a much harder product to build — but also a much more useful one if it works. (github.com)