AI Engineer World's Fair 2024
Personality-Driven Development: Exploring the Frontier of Agents with Attitude
Loading slides…
Read the talk
Personality-Driven Development: Exploring the Frontier of Agents with Attitude
Giving an agent a face changes how people use, price, and judge it, while prompts let its role and preferences become part of the business logic.
From a talk by Benjamin Stein
Giving software human traits
What changes when an AI agent gets a form and a personality? At Perpetual, Benjamin Stein builds agents presented as virtual teammates or AI employees. That framing draws on anthropomorphization: assigning human traits to nonhuman entities, for which Stein proposes the shorthand A18N. Yogi Bear, Lightning McQueen, and Nemo make the idea familiar before software enters the picture. He briefly contrasts it with assigning animal characteristics, which he calls zoomorphization, using Aslan as his example.
Software has used the same convention for decades. Clippy gives an assistant a recognizable character; the film Her imagines a relationship with a conversational operating system. Mailchimp’s monkey gives an email service a friendly representative. Today’s agents extend an established design choice: give an otherwise abstract capability something people can recognize and relate to.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
A voice can supply the persona
A character portrait is not required. Stein introduces Siri as an assistant without an explicit character, then asks the audience whether Siri—or Alexa—is female, male, or nonbinary, or whether the question itself is misguided. The exercise exposes how readily people assign human categories to a voice. He extends the question to GPT-4o: does the perception change when someone starts talking to it? The subject is the user’s attribution, not the system’s actual gender. Interface cues can create a persona even when the product has not explicitly designed one.
The consumer AI branding shown next moves in a different direction. Copilot, OpenAI, Meta AI, and Gemini use abstract representations rather than recognizable characters. Yet an abstract logo does not prevent users from supplying human characteristics once they interact with the system. Visual branding and perceived personality are not the same design decision.
The consequences can be surprisingly concrete. When Stein brought a Google Home into his kitchen, he demonstrated timers and music playback to his wife. She objected to the idea of bossing around the kitchen device. Explaining that it was not a person did not resolve the objection. Changing its voice did. The capability stayed the same; the voice changed the social meaning of using it in that household.
Google Photos offers the contrasting case: useful AI without a character. Stein describes searching for a dog and retrieving photographs of his beagle. The intelligence sits underneath an ordinary search interaction, with no personality to negotiate and no virtual employee to address. The visible result is simply the photos the user wanted.
Even that unobtrusive experience can acquire an identity in the user’s mind. When Stein showed it to his son, the child concluded that Google Photos had ChatGPT inside. The anecdote illustrates a brand association, not the product’s implementation: a familiar name becomes the explanation for an unfamiliar AI capability.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
A face creates expectations
Perpetual deliberately gives customers control over their agents’ forms, personalities, and preferences. The examples become increasingly unconventional, while the jobs remain recognizable:
- Cyborg tech lead: An android-like teammate writes code and performs code reviews.
- Artichoke recruiter: A vegetable with human characteristics represents a familiar workplace role.
- Hamster team: Characters occupy positions such as general manager and graphic designer.
The appearance can be arbitrary because the role supplies the organizing idea. A recruiter shaped like an artichoke is strange, but the word recruiter still tells the customer what kind of teammate they are looking at.
Once an agent has a face, customers assume they can converse with it. They ask whether they should contact it in Slack or Teams. That expectation is consequential because the underlying product is, in Stein’s description, smart workflow: tasks and operations that do not inherently require a chat interface. Personification creates an interaction requirement beyond the workflow itself.
Customers also expect a personality. The baseline is a helpful assistant: friendly, cooperative, and available to help. Perpetual lets customers vary that behavior with humor or snark, but the expectation comes first. A face seems to demand a manner of speaking, even when the underlying software could complete its work without one.
Failures then become personal. Stein describes customers cursing the hamster agents when things go wrong, contrasting that reaction with how people talk about failures in Google Sheets. The character gives frustration a target. For product development, the question becomes whether the benefits of personification justify the additional expectations and emotional responses it creates.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Roles explain the work and change the price anchor
The first benefit is explanatory efficiency. An AI software engineer immediately suggests a category of work. An AI recruiter suggests reading resumes and coordinating interviews, without a lengthy explanation of the system’s internals. A familiar job title communicates likely responsibilities in very few words. Giving the system a name also provides a conversational handle: people can discuss a recognizable entity more easily than an abstract algorithm.
Stein uses Google Now and Google Assistant to illustrate the naming effect. He recalls familiar tasks such as setting timers and reminders, but finds Assistant easier to understand because it describes something acting on a person’s behalf. This is a comparison about the meaning of the names, not a feature-by-feature equivalence between the products. Whether the Gemini name improves that clarity remains an open question in the talk.
The role can also change the buyer’s reference point for cost. Workflow software invites comparison with other subscriptions. A virtual employee invites comparison with staffing expense.
| Framing | Illustrative price comparison | Reference category |
|---|---|---|
| Workflow software | About $50 per month | Other workflow subscriptions |
| Junior employee | One-twentieth or one-hundredth of an employee’s cost | Staffing expense |
These are illustrative anchors, not Perpetual pricing or measured savings. Stein explicitly calls the workflow number made up and cautions that the broader pricing approach is not yet well tested. His early impression is that changing the reference category changes the executive conversation; whether that produces durable pricing power is unresolved.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Roles give agents boundaries
A role also provides a way to decompose an agent system. Engineering depends on choosing useful abstractions and dividing problems into manageable parts; job roles offer one such division. The boundaries are somewhat arbitrary, but they are understandable. Specialized agents can operate with a finite tool set and a smaller range of inputs. The rationale is that fewer choices leave fewer opportunities for an LLM to misinterpret a task or select the wrong tool. Stein presents this as a practical argument for specialization, without a benchmark comparison. As the problem grows, responsibilities can be subdivided further. An engineer writes code; a copywriter writes compelling copy. Each role makes a different part of the system easier to reason about.
There is also a separate benefit that does not require an accuracy or revenue claim: character creation is enjoyable. People spend time adjusting video-game avatars and creating Dungeons & Dragons characters because choosing an identity is part of the experience. Letting customers shape their virtual team can express a company’s brand and make the product more playful. Stein treats that enjoyment as a reason in its own right.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Personification creates product obligations
Generating characters for workplace roles introduces questions about representation and stereotypes. In Perpetual’s flow, customers choose a form and the product generates its appearance. Stein reports that 100% of the software engineer avatars generated in Perpetual’s workflow looked male and wore neckties. He attributes the pattern to DALL-E, but supplies no sample size, prompts, or model version. The observation concerns his generation workflow, not a universal output rate. Giving a role a visible body makes representational choices part of the product, raising the question of whether those choices are worth introducing in a work context.
The teammate metaphor also raises expectations of competence. Stein observes customers expecting agents to get every result right, even though they accept that junior employees make mistakes. Meanwhile, the expected chat interface consumes development effort whether or not conversation is central to completing the task. The character commits the product to both behavior and interaction features.
Those commitments may be difficult to reverse. Perpetual had not yet tried to remove its personas, but Stein anticipates that doing so would be much harder than changing colors during a rebrand. Customers would already have established characters and ways of working with them. Removing those identities would change their understanding of the product itself.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
The buyer and the employee hear different things
Personality, conversation, and preferences can draw attention away from the operational work customers actually pay for: document review, data extraction, and similar tasks. But Stein identifies a larger difficulty. Describing software as a virtual employee makes job replacement immediately salient. The label brings employment concerns into the conversation before anyone has established what the system will do or how it will affect the people using it.
A sales conversation shows how quickly the meaning can change. Stein pitched a CEO whose business could not hire enough people to grow as desired. The proposed agents would handle undesirable work at an attractive cost, and the CEO responded enthusiastically. At the first design session, the CEO brought an individual contributor who understood the actual work. He introduced Stein as someone with software that had virtual employees, then asked the employee: “Can you please tell him what your job is?”
The framing that appealed to the buyer made the employee conversation immediately difficult. Stein had to improvise a shift toward assistance and productivity: the software was supposed to help the person do their work. That reframing did not resolve what future employment outcomes would be. It exposed a product problem already present in the metaphor: the same virtual teammate can signify extra capacity to an executive and a possible replacement to the person asked to explain their job.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Preferences become business logic
The closing shift is from how an agent is represented to how its behavior is specified. Roles, job descriptions, forms, and personal preferences are supplied through prompts rather than a fixed set of checkboxes. Stein calls this zero configuration, but the configuration has not disappeared: natural language carries the behavioral specification. It lets customers inject nuance and business logic into their agents. Each instance can become bespoke, which complicates debugging. A failure under one customer’s instructions may not reproduce under another customer’s instructions, even when both are using the same product.
A constructed giraffe recruiter makes the layers concrete. Stein presents it as an example, not a demonstrated customer deployment.
| Layer | Recruiter example |
|---|---|
| Form | A giraffe |
| Personality | Charismatic and entertaining |
| Job description | Review resumes and read cover letters |
| Preferences | Favor Ivy League backgrounds, dislike frequent job changes, and value cover letters |
The preferences are illustrative hiring-manager inputs, not recommended screening criteria. They show where the specification moves beyond a general job description: the manager communicates priorities for how the work should be done.
This is training in Stein’s lowercase-t sense: telling the virtual employee how the manager wants it to work. In this prompt-driven setup, the example concerns instructions and preferences, not a demonstration of updating model weights.
The broader possibility is software molded to a business’s needs through language. Customers could define behavior beyond the checkboxes a SaaS product manager anticipated, making a description of how work should happen part of how the software is configured.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Resources
Further reading
Original 1994 experiments on social responses to computers, including voice and gender cues. Background for the talk's observations, not a source named by Stein.
Historical first-party account of Assistant's conversational framing and cross-device ambitions, useful alongside the Google Now naming comparison.
Historical documentation of image-generation bias and mitigation. It provides context without validating the talk's avatar percentage or establishing which DALL-E version Perpetual used.
Full transcript
Loading transcript…