Using Hermes Agent as a normal AI assistant is pretty straightforward.
You give it a task. It reasons about the problem, uses tools when necessary, and gives you a result.
But Hermes gets much more interesting when you ask a different question:
How do I get an AI agent to keep working without constantly prompting it?
Hermes Agent gives you several ways to do this.
You can create recurring loops, give the agent a persistent goal, schedule Cron Jobs, delegate work to sub-agents, create persistent specialist bots, coordinate multiple agents through Kanban, or even use multiple LLMs through Mixture of Agents.
The confusing part?
These features all increase agent autonomy, but they solve very different problems.
I recently published a video breaking down how these advanced Hermes Agent features work under the hood and, more importantly, when you should use each one.
🎥 Watch the full video:
Here’s the mental model I use to distinguish them.
🔁 1. Loop: Repeat Something Inside the Current Session
The easiest way to understand Hermes Agent’s /loop command is:
“Do this again later while I’m working in this session.”
A loop takes a prompt and reruns it on a recurring cadence.
For example:
Check my deployment every 5 minutes and tell me when it goes live.
The key distinction is that Loop is time-driven and session-scoped.
Use it for monitoring, repeated checks, polling, or other temporary recurring work associated with your current Hermes session.
🎯 2. Goal: Keep Working Until the Objective Is Complete
/goal solves a completely different problem.
Instead of saying:
“Do this again later.”
You’re saying:
“Keep working until this is actually done.”
Some tasks can’t reliably be completed in a single agent turn.
Maybe Hermes needs to modify code, run tests, inspect failures, make another change, and repeat the process.
That’s where Goal becomes useful.
After each continuation, the result can be evaluated to determine whether the objective has actually been completed.
Hermes also supports a completion contract, which can describe:
- The desired outcome
- How success should be verified
- Constraints that must remain true
- What is inside or outside the task’s scope
- When Hermes should stop and ask for help
Instead of:
Fix authentication.
You can define something closer to:
Fix authentication so all tests pass.
Keep the existing login API response unchanged.
Only modify the authentication service and its tests.
Stop and ask me if a database migration becomes necessary.
Now both execution and evaluation have a much clearer definition of “done.”
One important misconception:
Goal does not automatically create an agent team, Kanban workflow, or independent background process.
It keeps the current Hermes session focused on a standing objective.
So:
Loop → Do this again later.
Goal → Keep working until this is done.
⏰ 3. Cron Jobs: Scheduled, Unattended AI Agent Workflows
Both Loop and Goal are connected to an active Hermes session.
But what if you want Hermes to wake up every morning and perform some work even when you’re not interacting with it?
That’s where Cron Jobs come in.
For example:
Every morning at 9 AM:
- Check my news feeds
- Find important AI developments
- Summarize them
- Send me a report on Telegram
When the scheduled time arrives, Hermes can launch a fresh agent session specifically for that run.
This makes Cron useful for things like:
📊 Daily reports
🔍 Recurring research
🛡️ Security audits
🧪 CI health checks
📰 News monitoring
📨 Scheduled summaries
The fundamental distinction is:
Loop
↓
Keep checking while this session is active.
Cron
↓
Wake up independently on a schedule and run the task.
If I’m watching a deployment for the next 20 minutes, I’d probably use Loop.
If I want a report generated every morning whether I’m there or not, I’d use Cron.
🧩 4. Sub-Agent Delegation: Temporary Parallel Workers
Now we move from one autonomous agent to multiple agents working simultaneously.
Suppose I ask Hermes:
Compare three approaches for implementing authentication
and recommend the best one.
The main agent could investigate each approach sequentially.
Or it could delegate:
Parent Agent
│
├── Sub-Agent A → Research approach A
├── Sub-Agent B → Research approach B
└── Sub-Agent C → Research approach C
The sub-agents can perform those tasks independently and return their findings to the parent.
This provides two major advantages.
⚡ Parallelism
Independent work can happen simultaneously rather than forcing the main agent to perform everything sequentially.
🧱 Context Isolation
Large research, coding, refactoring, and analysis tasks can consume significant context.
Delegating them means much of the intermediate work stays inside the sub-agent’s context rather than flooding the parent’s conversation.
The important distinction is that delegation is temporary.
A sub-agent exists to perform a focused task, return its result, and finish.
If you want a specialist identity that persists over time, you need something else.
🤖 5. Profiles & Bot Mode: Persistent AI Specialists
Imagine having separate Hermes agents for:
👨💻 Coding
🔬 Research
✍️ Content
📈 Analysis
You probably don’t want your research agent’s accumulated context and workflows constantly mixed with your coding agent.
That’s what Hermes Profiles help solve.
A profile acts like a separate Hermes environment with its own configuration and state.
This allows specialist agents to develop independently.
Bot Mode builds a more intuitive multi-agent experience on top of these persistent profiles.
The key distinction from delegation is persistence:
Sub-Agent
↓
Temporary worker for a task
Bot / Profile
↓
Persistent specialist identity
Multiple bots can also communicate with one another, allowing you to create workflows involving specialist agents.
But once those workflows become complicated, messages alone aren’t enough.
📋 6. Kanban: Durable Multi-Agent Orchestration
Suppose you have three persistent Hermes specialists:
Researcher
↓
Analyst
↓
Writer
You want the researcher to collect evidence.
Only after the research is finished should the analyst evaluate it.
And only after the analysis is complete should the writer produce the final report.
You could coordinate everything through agent messages.
But what happens when the workflow introduces:
- Dependencies
- Failures
- Retries
- Artifacts
- Structured handoffs
- Human review
- Work lasting hours or days
Messages alone become fragile.
That’s where Hermes Kanban becomes useful.
Instead of the workflow existing only inside agent conversations, the work becomes structured, persistent task state.
Conceptually:
Research Topic
↓
Analyze Evidence
↓
Human Review
↓
Write Final Report
If Analyze Evidence depends on Research Topic, it waits.
Once the research task is completed, the dependent task can proceed.
This makes Kanban useful for durable, long-running multi-agent workflows.
And Kanban and delegation aren’t competing ideas.
A persistent Kanban worker could own a research task and then delegate parts of that research to several temporary sub-agents.
Think of it this way:
Delegation → temporary help
Bot Mode → persistent specialists
Kanban → coordination between specialists
🧠 7. Mixture of Agents: Multi-Model Reasoning
This is probably the easiest concept to confuse with multi-agent orchestration.
Mixture of Agents (MoA) operates at the model layer.
Imagine you’re solving a difficult software architecture problem.
Instead of asking one LLM to immediately decide what to do, you could ask several models to independently analyze the problem.
Reference Model A ──┐
│
Reference Model B ──┼──→ Aggregator Model → Hermes Agent Loop
│
Reference Model C ──┘
One model might notice a security problem.
Another might suggest a much simpler architecture.
Another might identify an assumption everyone else missed.
Their perspectives are then provided to an aggregator model, which becomes the model responsible for acting.
This is why Mixture of Agents is better understood as a model strategy rather than a traditional multi-agent workflow.
You’re not necessarily creating:
Research Agent
+
Coding Agent
+
Review Agent
Instead, you’re giving the acting model multiple model perspectives before it makes decisions.
The Hermes session, tools, and underlying agent loop can remain the same.
What changes is the model strategy powering that loop.
🗺️ The Mental Model
If you remember only one thing from this article, make it this:
| Hermes Feature | Mental Model | Best For |
|---|---|---|
| 🔁 Loop | Repetition | Repeated checks in the current session |
| 🎯 Goal | Completion | Working toward a clearly defined outcome |
| ⏰ Cron | Scheduling | Unattended scheduled workflows |
| 🧩 Delegation | Temporary parallelism | Independent subtasks |
| 🤖 Profiles / Bot Mode | Persistent identity | Long-lived specialist agents |
| 📋 Kanban | Durable orchestration | Dependencies, handoffs, retries, and review |
| 🧠 Mixture of Agents | Multi-model reasoning | Multiple LLM perspectives on difficult problems |
The features become much easier to reason about once you stop grouping all of them under “autonomous AI agents.”
They’re different primitives for controlling different dimensions of autonomy.
🚀 Putting It All Together
The interesting part is that these concepts can be combined.
Imagine a software development workflow where:
Cron
↓
Starts scheduled workflow
↓
Kanban coordinates persistent specialist profiles
↓
Research Agent delegates work to temporary sub-agents
↓
Architecture Agent uses Mixture of Agents
↓
Goal keeps implementation focused until verification passes
Now you’re moving beyond a chatbot that happens to call tools.
You’re designing an agentic system where scheduling, persistence, parallelism, identity, orchestration, and model reasoning are separate architectural decisions.
And that distinction becomes increasingly important as AI agents become more autonomous.
Top comments (1)
Curious how others are using Hermes Agent 👀
Which approach has been most useful in your workflows so far Loop, Goal, Delegation, Bot Mode, Kanban, or Mixture of Agents? And are there any combinations you’ve found particularly effective?