A named agent is a product. A generic assistant is a forgotten feature.
Give the agent a name, a role and a voice and you give the user something to form a relationship with, and a clear model of what it can do. Get the personality wrong and you over-promise capability. With a live demo.
See the live demoAn empty chat box is a hard thing to use. It offers everything and suggests nothing, so the user has to invent both the request and the expectation of what might come back. Naming the agent solves that in one move.
Give it a name, a role and a consistent tone, and the text box becomes a character. Not a gimmick, a model. The user now knows roughly what it is for, what to ask, and what good looks like. The name does the quiet work of setting expectations that a blank cursor never could.
The name carries the role
The reason a name works is that it is never just a name. "Finn, your accountant" or "Ace, your property manager" bundles a role, a domain and a promise of competence into one word. The user reads all of it instantly and calibrates accordingly. That is the upside, and it is large.
It is also the risk. A name that implies a capability the agent does not have is a broken promise waiting to happen. The anthropomorphism that makes a persona engaging is the same force that makes users over-attribute competence and inner life to it. Personality dialled up too far reads as a guarantee the agent cannot keep.
The craft is a persona that is warm and specific but candid about where it stops. Over-promise personality and you over-promise capability. The agents users trust most are the ones with a clear role and the nerve to say "that is outside what I do".
The same request, to two agents
One is a faceless assistant that offers to help with anything. The other is Zee, a named ops lead that knows its job and its limits. Toggle between them.
Can you sort out my overdue invoices and chase the late payers?
I’m an AI assistant and I’m here to help with a wide range of tasks. I can provide information and guidance. Could you tell me more about what you’d like to do?
A faceless feature. No role, no scope, no character. The user has no idea what it is actually for, so they get a vague offer to help with anything, which in practice means nothing. Nobody remembers this. Nobody comes back to it by name.
Note: Zee is p0stman’s own agent persona. Figures are illustrative.
Over-personification: the tells
A persona can tip from helpful into misleading. Watch for these:
Claims of feeling or caring
An agent that says it "feels excited" or "really cares" performs an inner life it does not have, priming the user to over-trust its judgement.
A name that out-promises the product
"Your lawyer" on an agent that cannot give legal advice. The name writes a cheque the capability cannot cash.
Confidence in place of competence
Personality used to paper over gaps. A charming agent that bluffs is more dangerous than a plain one that admits it does not know.
A persona that shifts between surfaces
Bubbly in chat, robotic in voice, anonymous on video. The user cannot tell it is one agent, so the relationship never forms.
What a good persona looks like
Six rules for a named agent that builds a relationship without writing cheques it can’t cash.
Give it a name, a role and a consistent tone
A name turns a text box into something a user can form a relationship with. A role tells them what it is for. A consistent tone makes it recognisable across every surface it appears on.
Let the name set capability expectations, then meet them
Calling it "your accountant" or "your ops lead" makes a promise about what it can do. The fastest way to break trust is to name a capability the agent does not actually have.
Be warm, but honest about the limits
The craft is a persona with personality that is still candid about what it cannot do. A friendly agent that admits its edges is trusted more than a slick one that pretends it has none.
Do not fake emotion or sentience
An agent that claims to "feel", "care" or "promise" is over-promising on the most dangerous axis. Personality is fine; manufactured inner life is anthropomorphism that sets the user up to over-trust.
One persona per product, consistent across channels
If the agent is Zee in chat, it is Zee in voice and Zee on the video call. A persona that changes voice between surfaces reads as several half-built features, not one product.
Degrade gracefully, in character
When the agent cannot help, the persona should hand off cleanly rather than bluff. "That’s outside what I do, here’s who can" protects the relationship far better than a confident wrong answer.
How to build it
Write the persona down before you write the prompt. A short character document, name, role, domain, tone, and crucially the explicit list of what it does not do, becomes the stable core of the system prompt. The limits matter as much as the personality: they are what stop the agent improvising past its competence.
Tie the name to the role, not to a mood. "Zee, your ops lead" sets a useful expectation. A cute name with no role attached gives the user nothing to calibrate against. The persona should answer the question "what is this for?" in its first line.
Keep one persona across every surface. The same name, voice and character appear in chat, in voice and on the video call, sharing the underlying conversation state. A consistent agent across channels reads as one product; a different voice per surface reads as several unfinished ones.
Bake graceful handoff into the character. When a request falls outside the role, the agent should say so in its own voice and route the user onward, rather than bluffing. Honesty delivered in persona protects the relationship better than competence faked in persona ever could.
Frequently asked questions
Don’t just read this. Put it to work.
The whole series is distilled into one Markdown file: every pattern, the do and don’t rules, and how well each is evidenced. Download it into your project, or paste the link into any chat with your agent and tell it to improve your agent UX. It’s free, no sign-up, no attribution required.
Use these Agent UX principles to review and improve our agent's interface: https://p0stman.com/agent-ux/agent-ux-principles.md
We build named agents, not nameless chatbots
Finn, Ace, Mo, Reena, Zee: every product we build is structured around a named agent with a clear role and honest limits, consistent across chat, voice and video. If you want an agent users remember and trust, that is the job we do.