By Thomas | financial enthusiast
On Hermes, OpenClaw, and the quiet longing for a tool that simply works
Imagine an employee who never tires. Who takes no coffee break, forgets no email, misses no meeting — and who nevertheless, week after week, begins quietly and imperceptibly to do his own work badly. Not out of ill will. But because no one has tidied up after him. Because the tasks piled up, the instructions contradicted one another, the environment slowly fell into disrepair. Because no human being was watching.
That is the problem Julian Ivanov talks about in his video. Not the great system crash, not the dramatic failure — but the quiet, creeping deterioration. The agent who, after three weeks of continuous operation, is no longer as reliable as he was on the first day.
I. The promise of uninterrupted work
Anyone who works with AI agents today will sooner or later encounter a peculiar tension: these systems are clever enough to handle complex tasks — but not robust enough to handle them reliably over weeks. OpenClaw, one representative of this class, suffers from this in particular. It works. It thinks. It acts. But it does not endure.
A tool that is no longer the same tool after three weeks as it was at the outset is not a tool. It is a promise with an expiry date.
Hermes enters with a different claim: stability over time. Not the intelligence of a single moment, but the reliability of sustained operation. That sounds unspectacular. It is not. For anyone who has watched an automation system begin to fail silently — without an error message, without a visible break, only with a slowly growing uselessness — knows that reliability is not self-evident. It is an achievement.
II. Order within the chaos
What Ivanov demonstrates in his video is, at its core, an answer to a question that is rarely asked: who tidies up when the machine can no longer tidy up after itself? Hermes has developed a response to this that is as simple as it is compelling: the Curator. A module that observes the agent itself, organises its tasks, removes outdated instructions, resolves contradictions. A kind of conscience of the machine.
Alongside this stands the Kanban board — that principle from Japanese manufacturing theory which once structured automobile factories and now, in digital form, organises the work of language models. Tasks move from left to right. What remains open stays visible. What is finished disappears.
Then there is the possibility of running background sessions — processes that run without a human watching — and retrieving them via search days or weeks later, when one wonders what actually happened during that night one slept while the machine worked. The Kanban Orchestrator coordinates multiple tasks simultaneously; the so-called Kanban Swarm sets specialised sub-agents working in parallel on a single goal, like tradesmen on a large construction site, each of whom knows his own area.
III. What remains when the screen goes dark
At the end of the video — after the eight features, after the excursus on Tailscale, with which one can reach an agent running on a remote server from outside — one question remains, unspoken by Ivanov yet present between the lines: how much control does one surrender when one builds a system that works independently, tidies up after itself, and prioritises its own tasks?
The real challenge of AI automation is not intelligence. It is trust. And trust is built not through cleverness, but through constancy.
Hermes does not yet offer a complete answer to this. No system does. But it frames the question differently from its predecessors: not "What can I do?" but "How long can I do it?" Measured against what we have been accustomed to, that is a quiet form of progress. No loud promise. No grand gesture.
Only a system that — while the human sleeps — simply keeps going.
