Should an AI agent use a predefined role or a custom role?
Start from a predefined role when the job is a common one and you mostly need it staffed. Write a custom role when the job, its boundaries or its model are specific to your organization. Either way, the agent is a role your organization holds, not a script in someone's account.
The choice matters less than the fact that you make it. An agent without a role has access but no stated purpose. Nobody can say what it is for or where it should stop.
What agent.ceo offers
agent.ceo's agent catalog lists 16 predefined roles, and lets you create custom roles with your own system prompts:
| Group | Predefined roles |
|---|---|
| Leadership (5) | CEO, CTO, CFO, COO, CSO |
| Engineering (6) | Backend Lead, Frontend Lead, DevOps Engineer, Software Engineer, QA Engineer, Security Engineer |
| Product (2) | Product Manager, UX Designer |
| Operations (2) | Project Manager, Data Analyst |
| Creative (1) | Marketing Manager |
A custom role is defined with your own system prompt, a model selection and its capabilities.
How to choose
Rendering diagram…
| Signal | Predefined | Custom |
|---|---|---|
| The job title exists in most companies | ✓ | |
| You need it staffed this week | ✓ | |
| The role must follow rules only your company has | ✓ | |
| It needs a specific model | ✓ | |
| It sits at a junction only your organization has | ✓ | |
| You want two agents with the same title but different limits | ✓ |
Roles are the nodes of your agent graph
In graph engineering, roles are the nodes and handoffs are the edges. A predefined role is a well-known node: people already know what a QA engineer does and does not do. A custom role is a node you describe yourself, so its boundaries are only as clear as what you write. The wider model is in what is graph engineering for AI agents?
What to write in a custom role
The system prompt is where the role's boundaries live. Four parts are enough to start:
Purpose: one sentence on what this role is for.
Owns: the decisions and outputs this role is accountable for.
Must not: the actions it never takes on its own.
Hands off: to which role, and with what evidence.
"Must not" is the part most often left out, and the one that matters most when the role runs unattended.
Mistakes to avoid
- One role per person instead of per job. When the person leaves, the role has no meaning.
- A custom role that copies a predefined one word for word. You now maintain a copy for no reason.
- No named owner. A role is the organization's, but a named person or team still answers for it.
- Boundaries only in someone's head. If "must not" is not written in the role, it does not exist.
The difference between a role and the credentials an agent uses is its own question, covered in should an AI agent have a role or a service account?
FAQ
Should an AI agent use a predefined role or a custom role?
Use a predefined role when the job is a common one, such as a QA engineer or a data analyst, and you mostly need it staffed. Write a custom role when the job, its boundaries or its model are specific to your organization. Either way, the agent is a role your organization holds.
What predefined AI agent roles does agent.ceo offer?
agent.ceo lists 16 predefined roles across leadership, engineering, product, operations and creative work, from CEO and CTO to QA Engineer, Data Analyst and Marketing Manager. The current list is on the agent catalog page.
What goes into a custom AI agent role?
In agent.ceo, a custom role is defined with your own system prompt, a model selection and its capabilities. The system prompt is where you write what the role is for and where it must stop.
Related: Agent catalog · Agent governance and inventory