Technical Context
I dug into the complaints about Hermes, and the picture is quite familiar: the agent is powerful, but in production, it's not the model that is frustrating, but the behavior surrounding it. When it comes to AI automation and proper AI implementation, it's these little things that end up burning the team's hours.
The first pain point I'd highlight is the compact or compress mode. Judging by discussions and FAQs, during long dialogues, Hermes begins to lose context quality, and compression is supposed to save the session. In practice, this doesn't always work: if the context has already "rotted," compression often just barely revives the agent rather than returning it to normal.
The second problem is less noticeable in demos but highly visible in actual work: skills feel unstable. Users complain that skills seem to drop off far too often, and self-improving starts to look not like learning, but like an endless shuffling of behavior. At this point, I completely understand the community's frustration.
The third issue is the most unpleasant for an engineer: I don't see any clear, built-in version control for skill changes in Hermes. If the agent "improves" itself in the wrong direction, you cannot transparently roll back the state. For a system that modifies its own behavior, this is a massive weak spot.
The typical workarounds are well-known: manually setting context_length, running /compress before degradation happens, starting a new session, or storing configs and skill definitions separately. But this is no longer the magic of an autonomous agent; it's manual AI integration with mandatory safety nets at every single step.
What This Means for Business and Automation
If I am building an agent for a real process rather than a flashy demo video, these things immediately impact the architecture. An unstable compact mode breaks long task chains. The lack of versioning in self-improving features kills trust in automated skill updates.
Who wins? Teams that are prepared to keep Hermes on a short leash: using an external git for skills, explicit checkpoints, a limited toolset, and strict session rules. Who loses? Those who want to turn on "self-learning" and just forget about maintenance.
At Nahornyi AI Lab, we usually resolve such issues at the architectural level: we move memory and skills into a strictly controlled layer, preventing the agent from randomly mutating everything. If you are facing a similar problem, instead of heroically fixing the chaos inside Hermes, you can calmly build an AI solution development process tailored to your workflow so that rollbacks, change control, and predictability become the standard, not a dream.