Agentic workflow: an AI agent that works 24/7 with a team of subagents

An agentic workflow is a process in which an AI agent plans the steps of a task itself, uses tools, checks the result and adjusts. Run that in a loop and the agent no longer waits for a question: it picks up the next piece of work itself. Choose a task, carry it out, have it checked, wrap it up and take the next one. Day and night. The agent does not do this alone. It splits the work across subagents that each have a role, such as exploring, building and reviewing. A stronger model steps in where the work gets hard.

This moves AI from a helper for one employee to a team that finishes work on its own. People set the goals and approve what matters. Below you will read how such an agent loop works and what it takes to run one safely around the clock. You will also see which work it can already do today and what it looks like in our own Hermes setup.

Dark ring with a cobalt-blue core, connected to five clusters of rounded blue, cream and gray blocks on a glowing loop

How does an agentic workflow differ from a chatbot or a cron job?

A chatbot answers when you ask something and then stops. A cron job or script runs at a fixed time and does exactly the same thing every time. An agent loop, which keeps an agentic workflow running, sits in between. A scheduler starts the agent at set times, but the agent decides what to do next, based on a task list and rules. It reads the task and chooses an approach. It uses tools such as a CMS, a terminal or an API, and checks the result before it closes the task.

The difference is judgment at every step. A script gets stuck as soon as the input differs from what it expects. An agent sees that a task is unclear, takes a different route or deliberately leaves the task open with a question for a person. You can read what an AI agent is and how it works in what is an AI agent.

Anthropic draws the same line in its December 2024 guide to building effective agents. It writes: "Workflows are systems where LLMs and tools are orchestrated through predefined code paths." Agents, on the other hand, "dynamically direct their own processes and tool usage." An agentic workflow puts that flexible agent inside a fixed frame of tasks and rules. Guides on agentic workflows often name three patterns: planning, tool use and reflection. You will find them in the five steps below: the agent plans, uses tools and has its own work checked.

How does a team of subagents work?

One agent that does everything itself fills up: every search, every file and every intermediate step stays in the same conversation. That is why a main agent splits up the work. It hands well-defined subtasks to subagents and only gets their final report back.

In Hermes this is called delegation. According to the documentation, each subagent starts with a completely fresh conversation and its own terminal, and it inherits the main agent's tools. Only the summary goes back to the main agent. By default, up to ten subagents run at the same time. You can also give subagents a different model. For example, a fast and inexpensive model does the research, while the main agent plans with a stronger one.

This creates a team with roles. An explorer gathers sources, files or data. A builder carries out a well-defined change. A reviewer checks the result without knowing the builder's reasoning. That is the point: a second pair of eyes that does not follow the same flawed line of thought. For demanding reviews, you put a stronger model on that role. A setup like this, with several agents working together, is also called a multi-agent system.

What does one round of an agentic workflow look like?

Every round follows the same five steps. The order is fixed; what the agent does within a step depends on the task.

  1. 01

    Choose a task

    The agent reads the task list and picks the next task according to fixed rules: urgent work first, then the oldest task. It skips tasks that a person has to decide on or that wait for a date. For us, that task list is a list of GitHub issues, but it can also be a ticketing system or a shared board. Hermes also has its own kanban board for multiple agents.

  2. 02

    Carry it out with subagents

    The main agent reads the full task, its history and the files that belong to it, and makes a plan. Research goes to fast subagents, a well-defined change to a builder. The main agent keeps the overview and decides.

  3. 03

    Have it checked

    Before the agent closes anything, a separate reviewer checks the work. Does it match the assignment? Are the sources real? Is the result live as intended? The reviewer approves or rejects, with evidence. After a rejection, the work goes back to be fixed.

  4. 04

    Wrap up and record

    After approval, the agent writes a report on the task: what changed, how it was checked and what was deliberately left out. That report is the memory for both people and the agent. Anyone who later wants to know why something is the way it is will find it there.

  5. 05

    Take the next one or stop

    Then the next round starts. If there is no task left that it can carry out, the agent stops without touching anything. Sometimes the reviewer rejects a task twice, or the task needs a decision that belongs to a person. Then the agent leaves it open with a specific question.

What do you need to run an agent safely 24/7?

An agent that keeps working at night also has to be able to stop without you. A good model alone is not enough for that. You agree on a few things up front:

  • A task list with clear assignments: one task is one well-defined result with criteria, so the agent knows when it is done.

  • Completion rules: the agent only closes a task after its own check and an approval from the reviewer.

  • A brake after repeated failure: Hermes counts how many runs of a scheduled task failed in a row and warns you after three by default. One command pauses all scheduled tasks at once.

  • Human approval where it matters: decisions about prices, claims, legal texts or anything that cannot be undone go to a person.

  • Limited permissions: the agent only gets the access the task needs, and read access where that is enough.

  • Logs and reports: every run leaves a trail that you can review afterward.

Hermes schedules this kind of work with scheduled tasks: you provide a schedule, an assignment and the skills the agent should load. You decide what goes into that assignment. You can read how to limit an agent's permissions in our practical safety checklist.

What does such a loop look like for us?

Our own Hermes setup at Voltti works on this website this way. The agent runs day and night. Each run picks one task from our task list and carries it out in the CMS or in the code. A separate reviewer checks the work, and the task only closes after that reviewer approves.

Other agents add to the list. A daily audit records improvements for each page as a new task, and another agent looks for new search opportunities every day. A separate watchdog pauses the loop when three runs in a row get stuck on the same authentication error. The loop never touches tasks that are marked as work for a person. We set the goals and the rules and review the reports.

The count for one week, from November 7 through November 13, 2026: the loop started 52 times, 50 runs finished and 2 failed. On November 10, we raised the schedule from six to nine runs a day. Within those runs, 32 tasks were completed and closed. In the same week, 48 new tasks were added, most of them created by the agents for audits and search opportunities. These are our own numbers for one website, not a benchmark.

Which work can an agent loop already do on its own today?

Work that consists of digital steps and has a result you can check. Think of content and SEO maintenance, well-defined code maintenance, reports and pulling data. The longer and vaguer a task, the bigger the chance that the agent takes a wrong turn along the way.

That boundary keeps moving. In March 2025, the research organization METR reported: "The length of tasks (measured by how long they take human professionals) that generalist frontier model agents can complete autonomously with 50% reliability has been doubling approximately every 7 months for the last 6 years." METR now notes that this figure is out of date and tracks a newer measurement. The direction stays the same: agents finish longer and longer tasks. The table shows where we see the boundary today.

Type of workWorks on its own today?Where a person stays involved

Content and SEO maintenance of a website

Yes, with a review and a report per task

Goals, claims and what may be published

Well-defined code maintenance with tests

Yes, for small and clear changes

Design choices and what goes live

Reports and pulling data

Yes

Interpretation and decisions

First line for customer questions

Partly, with fixed answers and handoffs

Exceptions, complaints and commitments

Quotes and prices

Preparing yes, sending no

Price and terms

HR, legal and financial decisions

No

Entirely

Where is a person still needed?

In everything that requires responsibility. An agent can update a legal text, but the business stays responsible for what it says. The direction is clear: more and more work in the workplace goes to agents. Still, 'replacing your entire workforce' is not a realistic promise today. What does change is the role of people: less doing it yourself, more steering, approving and keeping the rules sharp.

Also limit what an agent can break. Do not give admin rights where read access is enough, route irreversible actions through a person and keep the logs. Otherwise a mistake in a loop repeats itself on every run.

How does Voltti set this up with Hermes Agent?

Voltti helps businesses in Belgium and the Netherlands set up Hermes Agent, with a free choice of model and of API or subscription. For a loop like the one above, you start small: one task list and one type of work. Clear rules set what the agent may finish on its own. You keep skills, plugins and MCP connections in a GitHub repo of your own company. That way your team sees and manages how the agent works.

In a pilot, we test such a loop on one workflow. A pilot usually takes 2 to 4 weeks, from the first conversation to the evaluation; you can read how it works in our approach. Still unsure which workflow to pick first? Read why you should start with one frustrating workflow.

Portrait of Seppe Gadeyne

Seppe Gadeyne is a developer and the owner of Voltti. He sets up AI agents for companies in Belgium and the Netherlands.

Updated on

Frequently asked questions

01What is an agentic workflow?

A process in which an AI agent plans the steps of a task itself, uses tools, checks the result and adjusts. In an agent loop, the agent repeats this task after task, on a fixed schedule.

02What is the difference between an agent and a subagent?

The main agent plans and decides. A subagent gets one well-defined subtask from the main agent, completes it in its own fresh conversation and only returns a summary. This keeps the main agent focused and lets subtasks run at the same time.

03Can an AI agent really work 24/7?

Technically, yes: a scheduler starts the agent at set times, including at night. Whether that makes sense depends on the work. It only works safely with a task list, completion rules, a brake after repeated failure and human approval for decisions that matter.

04Who is responsible for what an agent does?

The business that uses the agent. The agent works within the permissions and rules you give it. That is why you decide up front what the agent may finish on its own and what goes through a person, and you review the reports.

05What does an agent loop cost?

That depends on the model, the number of runs per day and how many subagents take part. According to the Hermes documentation, subagents working in parallel typically use the large majority of the tokens. Voltti works with a custom quote. You can read which cost items play a role in what an AI agent costs.

An AI agent that picks up the next task itself?

Discuss your pilot