Video: "Hermes Agent: NEW Goals Update is Insane (FREE)!" by Julian Goldie on YouTube.
What the goals update actually does
Before this update, Hermes would complete one turn of a task and then pause. If you asked it to research competitors, produce a content brief, and draft three page outlines, it would do the first bit and stop. You re-prompted, it did the next bit and stopped again. Useful, but not exactly autonomous.
The /goal command changes the loop. You set a target — "build a complete SEO brief for this topic cluster" — and Hermes evaluates its own progress after every turn. If the goal is not yet met, it carries on. It only stops when it judges the work complete, or when it hits a genuine blocker it cannot resolve alone.
How the goal-checking loop works
The mechanism Nous Research call the "Ralph loop" is simpler than it sounds. At the end of each agent turn, Hermes checks the current state of the work against the goal definition. If there are still steps outstanding — a section not written, a file not created, a check not run — it queues the next action and continues. The loop exits when the success conditions are met.
This means you can give Hermes a reasonably complex brief — a 12-page SEO audit, a set of blog drafts, a Python script with tests — and walk away. It will handle sequencing internally, rather than treating each step as a separate one-shot prompt. Worth knowing: the quality of the goal description determines how reliably the loop terminates cleanly. A vague goal tends to produce vague self-assessment.
What kinds of work benefit most
The gain is clearest on jobs with a predictable sequence of steps. Content calendars, technical audits, website page builds — anything where "done" has a clear definition. The agent can work through fifteen linked tasks without you sitting at the keyboard to move it along. That is genuinely useful if you run a small team and need AI doing more of the groundwork overnight.
In practice it also reduces the cost of distraction. The main problem with current agent setups is that they require close supervision. Goal persistence moves the supervision from turn-by-turn to goal-definition-upfront, which is a better use of your time.
Where it still needs watching
Goal persistence is not the same as quality control. Hermes will complete the goal as it understands it — which may not match what you actually wanted if the initial description was ambiguous. The loop can also run longer than expected on complex tasks, burning through API credits if you are using a paid model. Running locally with an open-source model avoids this, but adds setup overhead.
Open-ended goals — "improve our website content" — should be broken into smaller, bounded targets before handing them to the agent. The more precisely you define done, the more useful this feature becomes.
Where this connects to NordSys
Getting AI agents to work through complex jobs reliably, rather than just completing individual prompts, is exactly the kind of design challenge we tackle for clients. The goal definition, the agent profile, the output format — these all need thinking through before a run. We help businesses set up Hermes and similar agents to work through real workflows without constant babysitting. Our AI Agents service covers the full setup.
See our AI Agents service →