In Chapter 1, you saw the problem: a chat tool that forgets everything between sessions and cannot act on its own. In Chapter 2, you saw why a persistent runtime matters. Now it is time to understand what lives inside that runtime.
An agent in Hermes is not just a model responding to a prompt. It is a complete working identity — a combination of a model, a system prompt, tools it can call, memory it carries forward, a working directory, and a configuration that ties it all together. When you define an agent, you are defining a role: who it is, what it knows, what it can do, and how it behaves.
Think of it like hiring for a position. You do not just hire "someone smart." You hire a specialist with a defined role, a set of responsibilities, access to specific tools, and instructions about how to do the work. Hermes agents work the same way. The model is the brain; everything else is the job description, the desk, the filing cabinet, and the schedule.
The rest of this chapter walks through each component that makes up a Hermes agent — and shows how they combine using your SEO workflow as the example.
A profile is an isolated identity. It holds its own system prompt, its own configuration, its own memory, its own skills, and its own session history. Two profiles on the same machine are two separate agents — they do not share context, they do not see each other's files, and they do not interfere with each other.
This matters because you want agents with defined roles, not generalists that try to do everything. A single profile with a vague prompt produces vague output. Separate profiles with narrow instructions produce sharper results.
In your SEO workflow, you need at least two distinct roles: someone to manage the overall content process, and someone to handle the research. The first role is your SEO manager. This profile holds the strategy context — your brand voice, your content calendar, your publishing standards. The second role is your research specialist. This profile focuses on finding keywords, analyzing search results, and returning structured data.
Later chapters will expand this team. But these two profiles are where the agent team starts: one directing the work, one doing the research.
A skill is a reusable procedure. It tells the agent how to carry out a specific type of work — step by step. Skills are stored as documents (in SKILL.md format) and loaded automatically when they are relevant to the task at hand. The agent does not load every skill at once; it picks the ones that match what it is doing.
Think of skills as standard operating procedures. A restaurant has a procedure for opening, a procedure for closing, and a procedure for handling a health inspection. The staff does not keep all three in their head at all times — they follow the one that applies to the current situation. Agent skills work the same way.
In your SEO workflow, the research specialist would have a skill for keyword research: which tools to call, how to filter results, what format to return the data in, and what to exclude. The SEO manager would have a skill for content briefs: how to turn keyword data into a structured brief, what sections to include, and what style rules to follow.
Over time, agents can also create and improve their own skills based on repeated work. If the research specialist notices that you always ask for the same filtering criteria, it can save those criteria into the skill so you do not have to repeat them. This is not magic — it is the agent following instructions to persist useful patterns. The agent stores what you tell it; it does not independently become smarter.
Memory is what separates a persistent agent from a disposable chat. In Hermes, memory means the agent stores facts, preferences, and context — and retrieves them automatically when they are relevant to the current session. When you close a session and start a new one, the memory is still there.
The built-in memory system uses two primary documents per profile: a general memory file for facts the agent has learned, and a user preferences file for your identity and style rules. Both are loaded into the agent's context at the start of every session. There are also optional external memory providers (like Honcho or Mem0) that add semantic search and richer recall — but the built-in system is enough to start.
In your SEO workflow, memory is where the compounding value lives. If you tell the SEO manager to always write in a specific tone, it stores that rule in memory. Next session, the rule is already in context. If the research specialist learns that a particular keyword source returns better results for your niche, it can save that finding. The longer you work with the agent, the less setup you need — because the agent carries forward what you taught it.
Memory is not perfect recall. The agent stores what you tell it to store and what it discovers during tasks. It does not remember everything from every session. You control what persists by giving clear instructions about what to save and what to discard.
A toolset is a group of related tools with shared dependencies. Hermes includes over 70 built-in tools organized into toolsets — file operations, web search, shell execution, and more. You enable or disable toolsets per profile, which means each agent only has access to the tools it needs for its role.
Toolsets are how an agent moves from generating text to actually doing things. A chat model can suggest that you search the web. A Hermes agent with the web search toolset enabled can perform the search itself, read the results, and act on them — all within the same session.
In your SEO workflow, the research specialist needs access to web search and file writing toolsets — so it can find keyword data and save results. The SEO manager needs access to file reading and writing toolsets — so it can read the research results, draft content, and save briefs. You would not give the research specialist the same toolsets as the manager, because the research specialist does not need to draft content. Narrow tool access keeps each agent focused.
The gateway is a single process that connects your Hermes agent to over 20 messaging platforms — Telegram, Discord, Slack, WhatsApp, Signal, Email, and more. Instead of opening a terminal window to talk to your agent, you send a message on the platform you already use. The gateway translates your message into an agent prompt, and the agent's response back into a message you can read.
This matters because it removes a barrier. If you have to open a terminal, navigate to the right directory, and start a session every time you want your agent to do something, you will use it less. The gateway lets you message your agent the same way you message a colleague — from your phone, from Slack, from Telegram.
In your SEO workflow, the gateway means you can ask your research specialist for keyword ideas from Slack, get the results in the same thread, and tell the SEO manager to draft a brief — all without leaving your messaging app. The agent is always reachable as long as the gateway is running.
Cron is the built-in scheduler. It lets you run agent tasks on a timed interval — every 30 minutes, every Monday at 9am, or any schedule you define. You can schedule tasks using natural language ("every Monday at 9am") or standard cron expressions. The agent runs the task, and the results get delivered wherever you want — back to your messaging platform, saved to a file, or chained into the next task.
Cron also supports a no-agent mode, where a script runs without any model involvement — useful for lightweight checks and data-fetching jobs that do not need AI reasoning.
In your SEO workflow, cron is how you go from asking your research specialist for keywords manually to having it run a keyword scan every Monday morning and deliver the results to your Slack channel. You do not need to remember to ask — the schedule handles the asking, and you handle the reviewing.
Subagents are temporary agents spawned for specific tasks. They start with fresh context (no prior session history), have access to the same tools as the parent profile, and return their results when done. Subagents are useful when you need parallel work — two research tasks at the same time, for example — without contaminating the main agent's context.
Kanban is the coordination system for multi-agent work. It provides a shared task board where agents claim tasks, report progress with heartbeats, and hand off completed work. If an agent stops responding (a "zombie"), the system detects it and can reclaim the task. This prevents the common failure mode of multi-agent setups: two agents trying to edit the same file, or a task sitting unclaimed because no agent picked it up.
In your SEO workflow, these concepts come into play once the team grows beyond two profiles. For now, the SEO manager directs the research specialist through direct task assignment — no kanban needed. But when you add more roles (a SERP analyst, a content writer, a reviewer), coordination becomes necessary. Later chapters will show how to design that coordination.
Here is how the components combine for your SEO workflow so far:
Holds the strategy context, your brand voice, and publishing standards. Uses file tools to draft briefs and content. Remembers your style rules across sessions.
Focused on keyword research and data gathering. Uses web search and file tools to find, filter, and save keyword data. Has a keyword research skill that defines the procedure.
Both profiles carry forward what they learned — the manager remembers your tone, the specialist remembers which sources work best for your niche.
You message the specialist on Slack to start a keyword task, and the results come back in the same thread. No terminal needed.
Later, you can schedule the specialist to run keyword scans every Monday and deliver results automatically. For now, you trigger the tasks manually.
The model is the engine inside each profile, but the profile, skills, memory, toolsets, gateway, and cron are what turn that engine into a useful worker. The model generates text; the surrounding components turn that text into action, persistence, and coordination.
By the end of this guide, your SEO team will grow from these two profiles into a coordinated group — with scheduled scans, shared memory, and publishing safeguards. But every addition builds on these same components. Understanding them now makes everything that follows easier to reason about.
Think about a recurring task in your own work. If you were to design two agent profiles for it, what would each profile's role be? What skills would each need? What memory would each carry forward?