Skip to main content
An agent is an AI coworker that someone in your organization sets up for one job, such as answering HR questions or checking invoices. You hand it a task in plain words, and it works on it until it can give you an answer.

What an agent is made of

Whoever builds an agent decides what it knows, what it may do and who may use it. The agent studio holds these choices in sections:
  • Identity: its name, its description and its type (an agent people talk to, or a subagent other agents hand work to).
  • Instructions: the job, the steps to follow and the tone to use.
  • AI model: the model the agent runs on.
  • What it knows: the knowledge bases it searches and the skills it can follow.
  • What it can do: tools such as web search, code execution or a browser, and toolkits that connect it to services like Jira or Outlook.
  • Subagents and Workflows: other agents it can hand part of a task to, and fixed processes it can start.
  • Guardrails: rules it keeps to, for example to cite its sources, to answer only from its knowledge bases, or to ask before certain actions.
  • Who can use it: private to its creator, shared with people or groups, or public to the organization.
People who use an agent get the version its builder last published. Changes wait in a draft until the next Publish.
  • Enterprise: Admins and Members build agents. Viewers use the agents shared with them, from New task.
  • SaaS: Admins and Members of a Business organization build and share agents. On a personal account (Free or Pro), your agents are yours alone and are never shared.

How it differs from a chatbot

A general chatbot answers from what its model learned. An agent in Neural Factory can do more than that, depending on how it is set up:
  • It searches your organization’s documents and shows which ones it used, so you can check its answer.
  • It follows your team’s way of doing a kind of work, written down as a skill.
  • It acts: it runs code, creates files you can download, browses the web, or works in a connected service.
  • It hands part of a task to a specialist agent (a subagent) and combines what comes back.
  • It stops to ask you a question, or to get your approval for an action its guardrails say needs one.
  • It keeps working on the platform when you close the browser tab, and it can run on a schedule.
An agent can still be wrong. Its sources are there so you can check.

An example

Your HR team sets up an “HR helper” agent. Its instructions say it answers staff questions about leave, benefits and expenses. Under What it knows, it has the knowledge base that holds the employee handbook and the leave policy. Under Guardrails, Cite sources is on, and Hand-off points people to the HR team for anything personal. You ask it: “How many days of annual leave do I get in my first year?” It searches the handbook, finds the passage on annual leave, and answers with a numbered marker. Clicking the number shows the passage it rests on. If the handbook does not cover your question, it tells you to contact HR.

Where you see this in the app

  • Workforce › Agents (Admins and Members): the list of agents you can see, and Build an agent to make one. See Agents overview.
  • New task: pick an agent and give it a task. See Selecting an agent.
  • Workforce › Catalogue (Admins and Members): ready-made use cases you can activate as a starting point. See Catalogue.

Next