← Research library

Agent personas are not characters: defining roles that hold up

AI-assisted research. This note was prepared with the assistance of AI tools. It may contain errors and is not professional advice. See our Legal Disclosures.

Agentic marketing programs now run several agents side by side, and teams reach for a familiar way to tell them apart: a name, a personality, sometimes an avatar. Personas make agents easier to talk about. Built around character instead of role, they also produce agents that sound confident and behave inconsistently.

"Meet Riley, our content strategist" is a friendly introduction, and there is nothing wrong with a friendly name. The trouble starts when the personality is doing the work that a job definition should do. In our research, what makes an agent dependable is a well-defined role: what it is for, what it touches, and how its work is judged.

This post covers the difference between the two, why character-heavy personas fail, and how to write a role definition you can test.

Character versus role

A character description says who the agent is like: "an experienced, witty marketer who loves data." That gives the model a voice to imitate. It gives little guidance about what to do, what to leave alone, or how to judge its own output.

A role description says what the agent is responsible for. We break it into six parts:

  • Purpose: the outcome the agent exists to produce.
  • Scope: the tasks it handles and the adjacent tasks it does not.
  • Inputs: what it receives and from whom.
  • Outputs: what it produces, in what format, and for which reader.
  • Standards: how its work will be judged.
  • Boundaries: what it must never do and when it must escalate.

Roles can be tested, because each part implies a check. Characters cannot, because there is nothing concrete to check against. If a reviewer cannot point to the line in the definition that an output broke, the definition is too vague to govern the agent.

Why character-heavy personas cause problems

Models take descriptions of identity seriously. An identity carries implied permissions that nobody wrote down, and the agent acts on them. We see four recurring problems.

Overconfidence. An agent told it is an "expert" may be less likely to admit uncertainty or ask for missing information. Experts, after all, are supposed to know. The result is gaps filled with plausible guesses instead of flags.

Scope creep. A strategist character may offer strategy when asked to draft copy, because that is what the character would do. Each extra is small, but reviewers end up cutting material they never asked for.

Inconsistent tone. A strong personality can leak into customer-facing content where it does not belong. Wit that suits an internal summary reads oddly in a renewal notice.

Misplaced trust. Team members may treat a named, personable agent as a colleague with judgment it does not have. People ask Riley for opinions they would never ask of "the draft generator," and they weigh the answers accordingly.

How the failure shows up in practice

Consider a hypothetical content team with an agent called Riley. Its persona, "a senior content strategist with a sharp eye for what converts," is the only description of its role. Riley is asked to draft three social posts announcing a webinar. The posts come back with a suggested change to the webinar's positioning, a new tagline, and a confident line about expected attendance that nobody supplied.

None of that was requested. Each addition is what a senior strategist might plausibly do, so the character licensed it. The team trims the posts, then trims the next batch, and the pattern repeats because nothing in the instructions marks the extras as out of scope.

You can spot this pattern in review notes before it becomes a habit. Look for these signals:

  • Unrequested additions: outputs contain recommendations or content outside the task.
  • Unsupported specifics: figures, dates, or claims appear that were not in the inputs.
  • Voice leakage: the persona's personality shows up in customer-facing copy.
  • Rare escalations: the agent seldom says it lacks information, even when inputs were thin.

Writing a good role definition

Use the language you would use in a clear job description for a contractor. Here is a role for a research agent:

You prepare account research briefs for the sales team. For each account, you summarize public business priorities, recent news, and relevant hiring, using only the sources provided. You do not recommend pricing or make claims about our product. If information is unavailable, you say so. Each brief follows the template and is under 300 words.

The definition says nothing about personality, expertise, or backstory. The tone of the output is governed by the brand voice specification, which every agent shares, instead of by a character each agent invents.

To write one for your own agent, work through the parts in order:

  1. State the outcome the agent exists to produce, in one sentence.
  2. List the tasks it handles, then list the adjacent tasks it does not.
  3. Name the inputs and who provides each one.
  4. Describe the output format and the person who reads it.
  5. Write the standards as checks a reviewer could apply.
  6. State the boundaries, including when to stop and ask.

Boundaries deserve the most care. An agent that knows when to stop is safer than one that knows a lot, and the stopping rules are the part character language erodes first. Our post on escalation paths covers how to write those triggers so the agent can recognize them.

Testing a role against a persona

A role definition earns its place when it changes behavior you can measure. Run the agent's golden dataset with the old persona and then with the new role, holding everything else constant. We look at four measures:

  • Scope adherence: the share of outputs that contain only what the task asked for.
  • Escalation rate: how often the agent flags missing information, and whether those flags were correct.
  • Edit rate: how much reviewers change before approving.
  • Format compliance: whether outputs follow the template every time.

Expect the escalation rate to rise at first. That is usually a good sign, because the agent is now reporting gaps it used to fill with confident guesses. Check that the flags are accurate before treating the rise as a problem.

The comparison also shows which parts of the role carry weight. If removing the boundaries section changes nothing on your test set, the test set probably lacks the cases that need them. Add a few inputs with missing facts or out-of-scope requests, and run it again.

Keep names if they help

Friendly names are fine for internal convenience. A short name is easier to say in a standup than a role title, and it helps a team talk about which agent produced what. The name is a label, and the role definition does the work.

Two cautions apply. Avoid presenting agents to customers as if they were people; see disclosure practices, and confirm the requirements for your markets with counsel. Inside the team, make sure the name does not stand in for permissions. An agent called Riley has exactly the access its role grants, which our post on roles and permissions treats as a separate, explicit decision.

The question we are studying next is how roles should change as an agent gains autonomy. A role written for drafting under review may need tighter boundaries, not looser ones, once the same agent can act without approval. We expect role definitions to become versioned documents that record each of those transitions.

How to apply this: rewrite one agent's persona as a role definition with purpose, scope, inputs, outputs, standards, and boundaries. Run its test set before and after, and compare scope adherence, escalation accuracy, and edit rate.

Related research

Put this research to work

Marketix turns these patterns into software: agents, automated workflows, and a hub that keeps people in control. Sign up for early access.