“Apps shouldn’t exist” - Karpathy

This is the repeated comment by Andrej Karpathy, Founding Member of OpenAI, former Director of AI at Tesla, and currently the founder of Eureka Labs.

What Andrej is seeing is a future where software apps that exist today have little value or relevance in the future where agentic AI does most of the work for us.

We are moving towards Agentic AI because we are now giving agency to the system itself

We have grown up in the age of software, where almost every interaction we have with digital systems is through a software interface. Our phone, our computers, service desks, even our televisions, are interfaces for software that we use to complete tasks.

We built these tools out of necessity so that we can leverage IT to achieve our objectives, and they have been very effective over these past decades.  But what if I told you the majority of the systems will disappear over the coming 20 years?

We use software because it gives us agency - it allows us to interact with organisations and other people extremely rapidly.  It allows us to engage large numbers of people at once.  It provides us the ability to complete tasks that previously were highly manual and time-consuming.

Karpathy argues that many current applications are "spurious" because they only exist to provide a human-centric interface that agents no longer require.

Karpathy says

"That app shouldn't exist... there's no need to have any of the app in between."

"A lot of this code shouldn't exist and it's just neural network doing most of the work."

"There's this sense that these apps that are in the app store... shouldn't even exist... shouldn't it just be APIs and shouldn't agents be just using it directly?"

"Everything should be a lot more just like exposed API endpoints and agents are the glue of the intelligence that actually like tool calls all the parts."

People believe that AI, or Agents, are just another layer in our software stack - “the user interface sits behind an agentic chat interface which in turn then sits on top of SaaS solutions that contain and manage data / task orchestration.”

This is missing the point.  We are moving towards Agentic AI because we are now giving agency to the system itself.  Except that the software apps we have built prevent the system from allowing that agency.  Our legacy systems are incapable of providing AI the ability to autonomously complete valuable objectives on behalf of human users.

Creating a chat interface over existing legacy systems doesn’t solve the problem of speed to completion, of personalisation, of agency and judgement.  It is simply putting an English-speaking support agent in front of a heavily convoluted software system.

These systems are built for the masses, and thereby force you to fit the mould of the general populace, and only through answering a plethora of questions, eventually narrow down to your specific scenario, so that it can help you specifically in your personal situation.

Let’s create a real world example to try and imagine what it will be like when Agentic AI tries to interact with our legacy software.

I am trying to do my small business taxes.  I have software apps that have many forms and fields and workflows and validations to help me do my taxes.  They force me to step through a pre-canned process and I’m required to answer many questions and provide lots of information to complete this task.  It still requires a lot of time, despite the many automations in the software and workflow.

The software is trying to fit me into a narrow bucket, or set of categories, so that it knows how to apply the correct rules and validations for my personal situation.  I’m trying to figure out how to answer some of its questions, because I don’t understand the background in sufficient detail.

The future is agentic, which is not a chat interface (a pre-canned support agent), no, it’s an intelligent agent that can navigate a system it has never seen before, and utilise tools and interactions for goal achievement.  Tools like create a report, submit a form, compile data, analyse results, and so on.

Let’s continue the example using an agentic AI Agent - one that can act on my behalf - and assume that these SaaS software apps still exist, and the agent must interact with them.

I ask the AI Agent to act on my behalf and perform the task for me.  Either it can directly interact with the software’s user interface, or it engages the software’s API (programmatic end point) that is built for software apps to interact more directly.  The software’s data entry and validations are still required and the AI Agent navigates the software, except it doesn’t have the contextual understanding when interacting with the software.  It must confirm what information it needs to provide to the software, and it must confirm my responses about me and my company’s activities, decisions, inputs, and results that it needs to be able to complete the forms.  

Eventually I will realise it would be easier if I just did it myself.

Now let’s imagine the same real world example, except with systems set up for Agentic AI.

I ask the Agentic AI to complete my taxes, but without using any tax software.  It studies the tax code relevant to my specific situation.  It identifies the information, format, and supporting evidence needed to complete the process.  It asks me to provide the information, and I clarify what documents, emails, receipts, etc. it needs from me.

The AI Agent takes all of this information, and compiles it into something that satisfies the reporting requirements of the tax office, including the format.  It’s careful not to simply trust it has done the job properly, so it consults one of the tax office’s verification agents that will confirm that the compiled information is correctly presented and it is complete.  

If there are corrections or inconsistencies or incomplete steps, then the verification agent will notify my Agentic AI, which will then continue to iterate on the task until it is completed to the verification agent’s satisfaction. 

The verification agent comprehensively implements all of the rules, validations, and controls that the original software app would have implemented.  Except the software is no longer built for humans to interact with.  It’s designed for agents to interact with.  Agents are able to deterministically complete and submit the taxes to the tax office’s requirements.

I never once had to deal with a software app.

This is the future.  The future is Agentic.

Karpathy says

"Everything has to be rewritten. Everything is still fundamentally written for humans and has to be moved around."

"I don't want to do anything - what is the thing I should copy paste to my agent... why are people still telling me what to do?"

"I'm hoping that there's a lot of agent-first infrastructure out there... [where] I could give a prompt to an LLM 'build menu gen' and then I didn't have to touch anything and it's deployed."

"The customer is not the human anymore it's like agents who are acting on behalf of humans and this refactoring will probably be substantial."

"I'll have my agent talk to your agent to figure out some of the details of our meetings."