# Personality Driven Development: Exploring the Frontier of Agents with Attitude Benjamin, Perpetual | AI Engineer World's Fair 2024 | 18:00 Source: https://www.youtube.com/watch?v=lG_8rgQqxfc Channel: AI Engineer (https://www.youtube.com/@aiDotEngineer). Summarised by AIE Talks. Page: https://aietalks.com/talks/personality-driven-development-exploring-the-frontier-of-agents-with-attitude Published: 2025-02-17 Tags: agents, design, prompt-engineering ## TL;DR - Giving an agent a face and personality makes customers treat it like a teammate, which changes what they expect from the product. - Personas give complex workflows an understandable name and provide a practical way to divide work among specialized agents. - Customer-specific prompts make every agent instance bespoke, which makes debugging, testing, and product support much harder. ## Summary Benjamin explains why Perpetual gives its AI agents names, forms, roles, and personalities. A face changes how users understand software. Customers expect to chat with the agent, assume it has a personality, and become more frustrated when it fails. Benjamin argues that personas also make complex workflows easier to describe. An AI recruiter or software engineer gives people an immediate model for what the system does. Roles can divide work into specialized agents with smaller tool sets and fewer inputs. The approach has serious costs. Generated avatars can reproduce gender and occupational stereotypes, personalities can distract from the paid business function, and a later rebrand would be difficult. The strongest concern is that virtual employees make job replacement visible during sales conversations. Perpetual's prompt-driven customization creates another engineering problem: each customer's agent is bespoke, so reproducing a bug or asking whether something works on one machine becomes difficult. ## Key ideas ### People assign human traits to software as soon as it has a voice or a face [00:33](https://www.youtube.com/watch?v=lG_8rgQqxfc&t=33s) Benjamin begins with anthropomorphization, giving human traits to non-human entities. He points to Clippy, Siri, Alexa, and other familiar products. Siri has no explicit persona, yet people still debate whether its voice sounds female, male, or something else. A Google Home conversation with his wife made the effect concrete. She objected to having a woman in the kitchen whom she could boss around, so Benjamin changed the voice to an Australian man. The software had not changed, but the perceived social relationship had. ### Perpetual makes agents feel like teammates by giving them roles, forms, and preferences [04:46](https://www.youtube.com/watch?v=lG_8rgQqxfc&t=286s) Perpetual builds what it calls virtual teammates or AI employees. Its agents have recognizable forms, job roles, personalities, and preferences. A tech lead can write code and review code. A recruiter can take a strange visual form, while teams of hamster characters handle roles such as general manager and graphic designer. Benjamin says the approach is deliberately unusual, but the role still gives users an understandable model of the work the agent performs. ### A visible persona makes customers expect conversation and a reliable personality [06:03](https://www.youtube.com/watch?v=lG_8rgQqxfc&t=363s) Once an agent has a form, customers quickly assume they can chat with it, perhaps in Slack or Teams, even when the underlying product is only workflow automation. They also expect a personality, usually the model of a friendly, helpful assistant. Failures feel more personal. People do not normally curse at Google Sheets when it breaks, Benjamin says, but customers curse at Perpetual's hamster agents. The face changes the user's emotional response to ordinary software failure. ### Agent personas give complicated workflows a name that people can understand [07:47](https://www.youtube.com/watch?v=lG_8rgQqxfc&t=467s) Benjamin says an AI software engineer or AI recruiter explains a product faster than a discussion of frameworks and tool calls. The role becomes a handle for talking about an otherwise abstract system. He compares this with Google Now becoming Google Assistant. The underlying functions were similar, but the word assistant made the system easier to understand as something that works on the user's behalf. He also suggests that a persona might affect price anchoring, since customers may compare an agent with the cost of a junior employee rather than with a software subscription. ### Roles provide a practical way to split work into specialized agents [10:24](https://www.youtube.com/watch?v=lG_8rgQqxfc&t=624s) Benjamin connects personas with engineering's usual work of choosing abstractions and decomposing problems. Specialized agents tend to have a finite set of tools and a smaller number of inputs, which gives the language model fewer opportunities to misinterpret a request or make an incorrect tool call. Roles also provide a natural way to keep dividing work into smaller areas. An AI engineer writes code, while an AI copywriter writes copy, so users can understand the division without learning the internal implementation. ### Personality creates social and product risks that the workflow itself does not have [12:16](https://www.youtube.com/watch?v=lG_8rgQqxfc&t=736s) Personified agents invite questions about inclusion and stereotypes. Benjamin says generated software-engineer avatars consistently appeared with neckties and masculine appearances. Customers also expect a higher level of performance from an agent that looks like an employee, even though human junior employees make mistakes. The team has to build chat interfaces and preference features because the persona makes them seem necessary. Benjamin also worries that the fun around characters can distract from the business value, such as document review and data extraction. ### Virtual employees make job replacement an immediate part of the sales conversation [14:26](https://www.youtube.com/watch?v=lG_8rgQqxfc&t=866s) Benjamin calls job replacement the biggest downside at the moment. He describes pitching a CEO on software that could help scale a business without hiring enough people. The CEO was enthusiastic until a later design meeting introduced the product as virtual employees to an individual contributor. That framing changed the conversation immediately and left Benjamin scrambling to explain that the system would help the employee. The story shows how a persona can make the social consequence of automation impossible to ignore. ### Prompt-driven customization makes every customer's agent a bespoke system [16:08](https://www.youtube.com/watch?v=lG_8rgQqxfc&t=968s) Perpetual lets customers define an agent's role, job description, form, personality, and preferences through prompts rather than fixed configuration controls. A hiring manager might tell a virtual recruiter to favor certain schools, dislike job hopping, or prefer cover letters. Benjamin is excited by the ability to mold software to a business's working style, but the engineering consequence is severe. Each instance is different, so reproducing a bug is difficult. The phrase "works on my machine" becomes almost meaningless when no two customer agents have the same configuration. ## Notable quotes - "As soon as you put a form onto something, they get real very quickly." (06:03) - "It is just a very stark reminder that you're replacing jobs." (14:26) - "Every single instance is 100% bespoke for each customer." (16:08) - "Of course this doesn't work, it's every single instance is bespoke." (16:46) ## Tools & references mentioned - Perpetual - Clippy - Siri - Alexa - Google Home - Google Photos - ChatGPT - DALL-E - Google Now - Google Assistant - Gemini - Slack - Microsoft Teams - Nintendo Switch - Dungeons & Dragons ## Who should watch - You are building agents for business workflows and need to decide whether a recognizable role or character will help customers understand the product. - Your product team is considering user-defined personas and needs to account for expectations about chat, reliability, stereotypes, and future rebranding. - Your engineering team is debugging prompt-configured agents where every customer's tools, preferences, and instructions differ. ## Related talks - [Designing Agents (The Floor Is the Frontier)](https://aietalks.com/talks/designing-agents-the-floor-is-the-frontier) (Ben Hylak, Raindrop, 19:46) - [Your AI Agent Isn't an Engineer: The Art of Thoughtful Anthropomorphism](https://aietalks.com/talks/your-ai-agent-isnt-an-engineer-the-art-of-thoughtful-anthropomorphism) (Rizel Scarlett, Block, 20:20) - [Form factors for your new AI coworkers](https://aietalks.com/talks/form-factors-for-your-new-ai-coworkers) (Craig Wattrus, Flatfile, 15:35) - [Rise of the AI Architect](https://aietalks.com/talks/rise-of-the-ai-architect) (Clay Bavor, Sierra & Alessio Fanelli, Latent Space, 18:55) - [Agents for Everything Else](https://aietalks.com/talks/agents-for-everything-else) (swyx, AI Engineer, 14:10)