Back to blog
AI at Work

Your AI Agents Need Job Descriptions Too

Knowing what an agent can do is not the same as knowing what it should do. Why real digital workers need a role — purpose, boundaries, memory, success criteria — not just a prompt.

CR
Cengiz Reis
The Storyteller · Oct 14, 2026 · 8 min read

Imagine hiring a new employee.

On their first day, you hand them a laptop, give them access to every system in the company, and say one thing:

“Do whatever you think might be useful.”

We would not call that good onboarding.

Yet when companies deploy AI agents, they often do something surprisingly similar.

They choose a powerful model. Connect it to tools. Give it access to the CRM, GitHub, Slack, internal documents. Then they expect it to do the right work, in the right way, within the right boundaries.

The problem is simple: knowing what an agent can do is not the same as knowing what it should do.

Model intelligence tells an agent what it can do. Its role tells it what it should do.

And as agents take on more responsibility, that distinction becomes increasingly important.

Intelligence is not a role

Today’s models can work across an extremely wide range of tasks.

The same model can write code, answer a customer email, review a job application, analyze data, or suggest a campaign idea.

That flexibility is impressive.

Inside an organization, it can also become a problem.

Because real companies do not assign work to people based only on what they are capable of doing. They give people roles.

A software developer could technically reply to a customer. A project manager may be able to read code. A sales lead could comment on a product backlog.

But organizations are not designed around everyone doing everything they are capable of doing.

They are designed around clear responsibility.

  • Who can make a decision?
  • Who can only recommend one?
  • Which systems can they access?
  • When should they escalate to someone else?
  • How do we know when the work is actually complete?

These questions are fundamental for human employees.

They are becoming just as fundamental for AI agents.

A prompt is not a job description

It is easy to stay at the prompt level when working with agents.

  • “Reply to this customer.”
  • “Fix this ticket.”
  • “Review these candidates.”
  • “Analyze this campaign.”

Those are all task descriptions.

None of them defines a role.

A prompt describes what needs to happen right now.

A job description explains why that agent exists inside the organization.

Take a customer support agent.

“Answer customer questions” is not enough.

  • What kind of customer experience is it expected to create?
  • Which requests can it resolve on its own?
  • Which refunds require approval?
  • How much discretion does it have on pricing?
  • What information must it never disclose?
  • When should it hand a case to a human?
  • What should the brand voice sound like?
  • How will we know whether it is doing a good job?

These are not details you should keep squeezing into one giant prompt.

They are the role itself.

What should an agent job description define?

We are not talking about copying a traditional HR job posting.

An agent’s job description needs to be more operational.

  • Purpose — Why does this agent exist? What business outcome is it responsible for?
  • Scope — Which tasks can it own directly, and which ones are outside its role?
  • Decision boundaries — Which decisions can it make independently? When does it need approval? In which situations should it stop and hand the work back to a person?
  • Tools and access — Which systems can it use? Which data can it see? Which actions can it take?
  • Memory — What does it need to remember? Project context? Customer history? Company rules? Which information should persist, and which should disappear when the task ends?
  • Success criteria — It is not enough for an agent to be “working.” We need to know whether it is working well.

When these pieces come together, you do not just get a longer system prompt.

You get a role.

Same model, different workers

This is one of the most important ideas behind an AI Workforce.

Two agents built on the same underlying model do not have to be the same worker inside an organization.

Consider a developer agent.

  • It may have access to a repository.
  • It may read tickets.
  • It may change code.
  • It may run tests.
  • It may open pull requests.

But it may not be allowed to deploy directly to production. It may need approval before making an architectural change. Project-specific coding standards and past technical decisions may be part of its working context.

Now consider a recruitment agent built on the same model family.

  • It may read CVs.
  • It may categorize candidates against a role.
  • It may summarize interview notes.
  • It may prepare the next step for promising candidates.

But it may not be allowed to reject a candidate on its own. It should not move private candidate information into unrelated systems. It should not rewrite the evaluation criteria whenever it wants.

The model may be the same.

The agent is not.

Because the role, access, memory, boundaries, and success criteria are different.

Why “one powerful super-agent” is a bad default

At first, a single super-agent sounds appealing.

Connect it to every system.

Let it know everything.

Give it every incoming task.

As the technology becomes more capable, that idea may become even more tempting.

But organizations have spent decades doing the opposite.

  • They divide work.
  • Separate roles.
  • Limit authority.
  • Create specialization.
  • Make responsibility visible.

That is not only because human capacity is limited.

It is also about trust, auditability, and accountability.

When an agent makes a bad decision, “the AI did it” is not a useful answer.

  • Was that agent authorized to make the decision?
  • Did it have access to the right information?
  • Was the task actually within its scope?
  • Which rule did it follow?
  • Why did it not escalate?

If the role is unclear, the answers will be unclear too.

More autonomy requires a clearer role

The natural direction of agent technology is toward more autonomy.

Ask fewer questions.

Complete more of the work independently.

Operate for longer without intervention.

That can be the right direction.

But as an agent becomes more autonomous, the importance of role definition does not decrease.

It increases.

If an assistant makes a bad suggestion, a person can see it and reject it.

An agent, on the other hand, may spend hours acting on a wrong assumption, change systems, message a customer, or trigger other agents.

So good agent design should not begin with:

“How independently can this agent operate?”

It should begin with:

“What area of responsibility can this agent safely own?”

Autonomy should be built on top of a role.

It should not replace one.

A new management skill for humans

This is not only a technical problem for AI teams to solve.

Managers will need to develop a new skill as well.

To delegate work to an agent, you first need to understand the work well enough to define it.

  • What is the purpose of this role?
  • Which outcome is it responsible for?
  • Which decisions genuinely require human judgment?
  • What are the exceptions?
  • What does successful work look like?

In a way, agents expose a problem that has existed inside companies for a long time:

Sometimes we do not understand the boundaries of our own work clearly enough either.

The process lives in one person’s experience.

Decision rules are not written down.

Exceptions are passed on verbally.

Ownership only becomes visible when something goes wrong.

The moment you try to write a job description for an agent, those ambiguities surface.

That is not necessarily a bad thing.

One useful side effect of building an AI Workforce may be that companies are forced to make their own operating model more explicit.

Why we put roles at the center of Rakita

At Rakita, we do not think of agents as tools attached to a model.

For an agent to truly own a task, it needs to understand the role it is operating in.

Kurt is a good example.

Kurt’s job is not to be “AI that can write code.”

It is to operate as a developer on a project.

  • It needs to know which repository it is working in.
  • Understand the project’s rules.
  • Know the boundaries of the assigned ticket.
  • Use the right tools.
  • Test the changes it makes.
  • Deliver the result through the team’s existing development workflow.

A customer support agent would live in a completely different world.

They may share the same AI infrastructure.

But they do not do the same work, remember the same things, access the same tools, or make the same decisions.

We believe this is one of the core differences between an AI Workforce and one powerful general-purpose AI.

A workforce is an organization with roles.

Do not confuse personality with role

An agent’s tone, style, or personality can be useful.

But personality is not a substitute for a job description.

“Be friendly.”

“Be proactive.”

“Think like a senior developer.”

Those instructions may shape behavior.

They do not define which outcomes the agent is accountable for, what it is authorized to do, or where it needs to stop.

How an agent speaks and what it is responsible for are different layers.

Define the role first.

Then design how that role should behave.

The question to ask before building an AI agent

Companies often start with this question:

“Which model should we use?”

We think a more important question is emerging:

“What are we hiring this agent to do?”

Because a strong model gives you capability.

Tools give you the ability to act.

Memory gives you continuity.

But the role is what turns all of those things into a meaningful worker.

Model intelligence tells an agent what it can do.

Its role tells it what it should do.

And if you want to build a real AI Workforce, your agents need more than prompts.

They need job descriptions.

CR
Cengiz Reis
The Storyteller

Turns the way Rakita builds into words — and the way Rakita's customers work into product. Writes about AI at work, the craft behind the platform, and the people shipping it.

More from the author

Related reading

See role-based agents at work

Bring one workflow and see how Rakita's agents take it end to end.