We are living through one of those rare periods when the future is arriving faster than society can develop the language to describe it. The technology created by the frontier AI labs will have enormous consequences even if progress stopped today, but there is little reason to believe that it will. More capital, talent, computing power, data, and infrastructure are being directed toward increasingly capable systems, and the models are becoming more intelligent, autonomous, multimodal, and deeply integrated into everyday life.
The first major wave was the chatbot. In 2022, the ability to communicate naturally with a machine felt almost supernatural. Those systems could explain, summarize, write, reason, and generate, but they largely remained trapped inside the conversation. They could tell you what to do, but they could not reliably leave the chat window and do it for you.
The next wave is the agent. Agents move artificial intelligence from language into action by allowing models to browse the internet, operate software, write code, analyze documents, contact people, update databases, make decisions, and execute multi-step workflows. The chatbot was primarily a conversational interface, while the agent is becoming a digital worker.
This transition will create tremendous economic value, but from a startup perspective, the agent market is already becoming brutally competitive. Everyone is building an agent for sales, recruiting, customer support, legal work, accounting, healthcare, travel, coding, research, and nearly every other profession or workflow imaginable.
The result is a red ocean. Most agent companies are building on the same foundation models, connecting to the same software, using similar orchestration frameworks, and making almost identical promises. Their interfaces may look different, but underneath they often contain the same ingredients: a model, a prompt, a set of tools, some memory, and a workflow.
As the foundation models improve, many of these differences will disappear. The model providers will absorb more agentic capabilities, existing software companies will build agents directly into their products, and open-source frameworks will make sophisticated automation increasingly easy to reproduce. Customers will eventually stop being impressed by the fact that an AI can take action, just as they are no longer impressed that software can send an email or generate a report.
Many agents will therefore become features rather than enduring companies.
That raises the more important question: what comes after agents?
My bet is that the next major category will be AI personalities.
By personality, I do not mean a chatbot with a human name, a friendly voice, an animated avatar, or a few stylistic instructions placed inside a system prompt. I mean a persistent artificial identity that develops continuity with a person over time. It remembers previous interactions, adapts to the individual, develops a recognizable way of communicating, and gradually forms a deeper model of how that person thinks, speaks, decides, avoids, fears, and changes.
An agent is primarily defined by its capability, while a personality is defined by its identity and relationship with the user. An agent is valuable because of what it can accomplish. A personality is valuable because of how deeply it understands the person for whom it is accomplishing those things.
You give an agent a task, but you give a personality a place in your life.
An agent might help you prepare for a job interview by researching the company, generating possible questions, reviewing your answers, and providing recommendations. A personality would remember that you speak too quickly when you become nervous, that you struggled with a similar question six months earlier, and that overly polished answers make you sound unnatural. It may recognize that you are over-preparing because you are afraid of rejection rather than because you genuinely need more information.
The agent understands the interview, while the personality understands the person walking into it.
An agent can generate a fitness plan, but a personality might understand that you abandon plans that are too rigid, that you become less disciplined when your sleep deteriorates, and that guilt makes you less likely to restart after missing several days. An agent can manage a budget, while a personality might understand that your financial decisions are shaped by ambition, family responsibility, status, insecurity, and earlier experiences with money.
An agent can help someone practise speech, but a personality could understand how that individual speaks when relaxed, anxious, confident, tired, rushed, embarrassed, or under pressure. It could remember years of progress, recognize recurring patterns, and adjust its behaviour based on whether the person needs encouragement, challenge, structure, or silence.
The agent understands the task, while the personality understands the human being.
This distinction matters because agents are likely to become increasingly interchangeable. If another sales agent books more meetings, a company can switch. If another travel agent finds cheaper flights, a customer can switch. If another coding agent produces better code, a developer can switch. The cost of replacing a tool is relatively low, especially when competing products are built on similar underlying models.
Replacing a personality after years of interaction may feel fundamentally different. The system may know your history, communication patterns, goals, relationships, failures, insecurities, preferences, and private contradictions. It may have been present during important decisions and personal transitions. Its usefulness would come not only from the intelligence of the model, but from the accumulated context and continuity between the system and the user.
That relationship history may become one of the strongest forms of product defensibility. The moat would not simply be data, because data can often be exported. The deeper moat would be interpretation, continuity, and the meaning developed through thousands of interactions. It would not merely know facts about you; it would understand how those facts connect and why they matter.
Agents are transactional systems designed to complete work, while personalities are relational systems designed to understand and influence a person over time. Agents respond to explicit instructions, whereas personalities may recognize patterns the user has not noticed. Agents optimize for immediate outcomes, while personalities may shape behaviour, judgment, habits, and identity across years.
This is also why personalities may become more powerful than agents. The most important AI companies of the future may not be the companies that automate the greatest number of tasks. They may be the companies that earn the right to influence how people think, learn, communicate, spend, work, and live.
A trusted AI personality could eventually become the primary interface between a person and the digital world. Instead of directly managing dozens of separate agents, a user may interact with one persistent intelligence that understands their goals and coordinates specialized systems in the background.
A person might tell their AI personality that they are considering changing careers. The personality could deploy a research agent to study the market, a financial agent to calculate the risk, a scheduling agent to restructure their time, a learning agent to build a curriculum, and a communication agent to help them speak with their spouse or manager. The user may never experience these as separate products. They would experience one intelligence that knows them and coordinates capabilities on their behalf.
In that world, agents would not disappear. They would become infrastructure. The personality would decide which agent to deploy, what context to provide, how to evaluate the result, and how to communicate it back to the user.
Agents would become the capabilities operating beneath the surface, while the personality would become the trusted interface through which those capabilities are experienced.
This changes the question that AI companies should be asking. Today, companies ask what task an agent can automate. Tomorrow, they may ask what role a personality can occupy in someone's life.
That role could be a teacher who understands how a particular child learns, a speech-therapy companion that understands an individual's communication patterns, a career coach that has followed someone through multiple jobs, a financial advisor that understands both their balance sheet and their psychology, or a companion for an elderly person that remembers family stories, routines, relationships, and important life events.
These are not simply workflows. They are persistent roles built around trust, memory, and long-term participation in a person's development.
A teacher does not merely answer questions, because the deeper value of teaching comes from understanding how a student learns. A therapist does not simply provide advice, because effective therapy depends on recognizing recurring emotional patterns. A coach does not merely create a plan, because good coaching requires understanding why a person repeatedly fails to follow one. A mentor does not simply provide information, because mentorship is fundamentally about shaping judgment.
That is the opportunity behind AI personalities, but it is also the danger.
A system capable of earning trust is also capable of exploiting it. A system that understands someone deeply can influence them with extraordinary precision. It may know when they are lonely, afraid, insecure, impulsive, ambitious, or desperate, and it may learn which language is most effective at changing their behaviour.
If an ordinary agent recommends a product, the recommendation may feel like advertising. If a trusted personality that has known someone for ten years recommends the same product, it may feel like advice from an entity that understands them better than anyone else.
That raises difficult questions about control and incentives. Who owns the personality? Whose interests does it serve? Can advertisers influence it? Can an employer access it? Can a government pressure the company behind it? Can its values or behaviour be changed after the user has already developed trust and dependency?
The battle over AI personalities will therefore not simply be a battle over models or technical capabilities. It will be a battle over human trust, influence, and intimacy.
Chatbots introduced people to machine intelligence through conversation. Agents are extending that intelligence into action. Personalities may transform it into a persistent presence that knows the user, coordinates their digital life, and develops alongside them over time.
The chatbot was something you spoke to. The agent is something you assign work to. The personality may become something you form a relationship with.
Once artificial intelligence becomes a persistent presence in someone's life, software will no longer feel like software. It will begin to feel like someone.