Organizations would never hire a colleague by giving them one instruction and unlimited access. Yet AI agents are often deployed through a prompt that describes a task while leaving authority, quality and escalation implicit.
If an agent performs recurring work, it has effectively acquired a role. That role deserves the same organizational clarity as human work.
A prompt is not a role
A prompt explains what to do in one interaction. A role explains why the work exists, how success is judged, which resources may be used and where authority ends.
Without this context, an agent may optimize the visible task while undermining the broader objective.
Elements of an agent job description
A robust definition includes:
- Identity and purpose: the organizational role the agent represents.
- Primary directive: the outcome it is expected to create.
- Scope: tasks and domains that belong to the role.
- Knowledge: approved sources and their precedence.
- Tools: actions the agent can perform.
- Mandate: what it may decide without approval.
- Quality standards: accuracy, format, evidence and timeliness.
- Prohibited behavior: actions or reasoning patterns it must avoid.
- Escalation: conditions that require another agent or a person.
- Accountability: the human owner responsible for performance.
Separate capability from permission
A model may be technically capable of sending messages or changing records. That does not mean the agent should be authorized to do so. Least privilege keeps the operational mandate narrower than technical possibility.
Permission can expand when evidence shows that the agent performs reliably and the organization can manage exceptions.
Manage performance as a system
Agent quality depends on instructions, knowledge, tools and feedback. Monitoring only the final answer misses the cause of failure. Organizations should review source use, uncertainty, human corrections and downstream outcomes.
Convene AI OS gives agents profiles, knowledge scopes, skills, permissions, tasks and auditable activity. This makes the role visible and governable.
The better question is not whether an agent is intelligent. It is whether the organization has designed a role in which that intelligence can be useful and accountable.