Garry Tan Spends Rs 1.8 Crore a Year on AI Agents

Artificial Intelligence is rapidly moving beyond answering questions and generating content. For Y Combinator CEO Garry Tan, AI agents have become an integral part of his personal and professional routine, helping him organise information, work on coding projects and prepare for important conversations. The scale of his AI usage is reflected in his spending: Tan says he spends as much as $200,000 (approximately Rs 1.8 crore) annually on AI tokens.

His approach offers a glimpse into a possible future where people rely on personalised AI systems that understand their preferences, professional responsibilities and personal history. Unlike conventional chatbots that respond primarily to information provided during individual conversations, Tan’s AI setup has access to an extensive collection of his existing digital records.

Speaking to author Walter Isaacson on his podcast, The Builders, Tan explained that his AI agents can access his entire email history, Dropbox files and Apple Notes. The systems have also reviewed his therapy notes spanning approximately 20 years.

This extensive access allows the agents to provide assistance using information about his life that would otherwise require considerable time and effort to explain to a conventional chatbot. However, it also raises important questions about privacy, data security, personal autonomy and the cost of building AI assistants with such extensive access.

Why Y Combinator CEO Garry Tan Spends Rs 1.8 Crore on AI Tokens

Tan’s annual AI expenditure of up to $200,000 demonstrates how expensive intensive AI-agent usage can become when systems are continuously processing information and assisting with multiple tasks.

AI tokens are units used by artificial intelligence models to process and generate information. They represent portions of text or other input and output data that models handle during interactions. Depending on the model and service, token usage contributes directly to the cost of operating AI applications.

When someone uses an ordinary chatbot occasionally, token consumption may remain relatively limited. However, AI agents can operate across multiple tasks, examine large amounts of information, interact with files and repeatedly process instructions to complete complex assignments.

This creates a different cost structure from traditional chatbot usage.

Tan’s setup involves AI agents that can draw on extensive personal information and assist with activities throughout his day. Their usefulness depends partly on their ability to understand the context behind his requests rather than treating every interaction as an isolated conversation.

For example, preparing for an important professional discussion may require understanding previous correspondence, relevant documents and the background of the people involved. An AI agent with access to these materials can potentially bring together relevant information without requiring the user to manually search through different applications.

The more frequently such systems operate and the more information they process, the greater their potential token consumption becomes.

AI Agents Have Access to Garry Tan’s Emails, Files and Therapy Notes

One of the most unusual aspects of Tan’s AI arrangement is the depth of personal information available to his agents.

According to his account, the systems have access to:

• Email history: His complete email archive, providing context about professional relationships, previous conversations and ongoing responsibilities.
• Dropbox files: Documents and other stored information that can help the agents understand his projects and work.
• Apple Notes: Personal notes that may contain ideas, reminders and information collected over time.
• Therapy notes: Records covering approximately 20 years, giving the systems access to highly personal historical information.

This level of integration changes how an AI assistant can interact with its user. Instead of depending entirely on what someone remembers to mention in a prompt, the system can potentially use previously stored information to understand the situation.

Tan believes this makes AI agents useful across almost every part of his daily routine. They can offer suggestions, assist with decisions and anticipate what he might need based on the information available to them.

He compared the experience to having a constant personal adviser who understands the user and remains available whenever assistance is required.

Describing the concept, Tan said: “Think of it as a Jiminy Cricket on your shoulder at all times. Like, almost anything you want to do, it knows you, and it’s there to help.”

The comparison captures the difference between a traditional digital assistant and a more personalised AI agent. The latter is not simply waiting for instructions but is designed to use accumulated context to make its assistance more relevant.

How Personalised AI Assistants Could Change Everyday Life

Tan’s experience points towards a broader shift in how people may interact with artificial intelligence. The next stage of AI adoption may involve systems that understand a person’s ongoing responsibilities rather than responding only to individual requests.

A conventional chatbot usually needs users to provide relevant background information whenever they begin a new task. A personalised AI agent, by contrast, can potentially retrieve that information from connected accounts and previously stored documents.

This could make several everyday activities more efficient.

Managing Professional Responsibilities

For professionals handling multiple projects, an AI agent with access to emails, documents and notes could help organise information and prepare for important discussions.

Instead of manually reviewing an entire email conversation, a user could ask the agent to identify relevant developments, summarise previous discussions or prepare background information before a meeting.

Such systems could also help users maintain continuity between projects by remembering earlier decisions and identifying information that may otherwise be overlooked.

Preparing for Important Conversations

Communication often requires understanding the history of a relationship or discussion. An AI assistant with access to previous correspondence could help prepare talking points, identify unresolved issues and suggest relevant questions.

For someone managing a large professional network, this type of assistance could reduce the time spent searching through old messages and documents.

Supporting Personal Organisation

Personal notes, reminders and historical records can contain useful information that people may struggle to recall immediately.

By connecting these sources, an AI agent could potentially help users organise their schedules, retrieve previous ideas and make decisions using information already available in their digital records.

However, these capabilities depend on the accuracy of the underlying information, the quality of the AI model and the permissions granted to the system. Access to more data does not automatically guarantee better decisions.

Garry Tan’s AI Obsession Has Even Affected His Sleep

Tan’s enthusiasm for artificial intelligence extends beyond his professional responsibilities. He has previously joked about experiencing “cyber psychosis” and barely sleeping because he becomes so excited about experimenting with AI agents.

His comments illustrate how quickly the Technology has captured the attention of people working closely with advanced AI systems.

For developers and technology leaders, AI agents offer an opportunity to experiment with new ways of working. Tasks that previously required several manual steps can increasingly be delegated to systems capable of processing information and carrying out complex instructions.

This can make experimentation particularly engaging, especially when an agent successfully completes a task that would otherwise consume considerable time.

However, the enthusiasm also raises a practical question about maintaining boundaries between work and personal life.

When AI tools are continuously available and capable of handling increasingly complicated assignments, users may feel encouraged to keep experimenting rather than disconnecting. Tan’s remarks about his sleeping habits offer a personal example of how enthusiasm for emerging technology can affect everyday routines.

Y Combinator Has Its Own AI Agent Called Stray Light

Tan’s interest in AI agents is not limited to his personal setup. The technology is also being used within Y Combinator, the startup accelerator he leads.

The organisation has an internal AI agent called Stray Light, which has access to information about founders funded by the accelerator, people who have contacted the organisation through email and startup applications.

According to Tan, the agent also has its own personality, making it more than a basic information retrieval system.

The system’s access to organisational information could help Y Combinator work with the large volume of information generated through startup applications, founder communications and its investment activities.

For an accelerator that evaluates startups and maintains relationships with founders, the ability to retrieve relevant information quickly can be valuable.

For example, an internal agent with access to application information and previous communications could potentially help employees locate relevant records and understand the background of a particular founder or company.

However, the existence of Stray Light does not mean that every organisational decision is being handed over to AI. The available information establishes that the agent has access to company-related data, but does not specify that it independently makes investment decisions or replaces human judgement.

Garry Tan Publishes 23 Claude Code Skills Under gstack on GitHub

Tan has also shared his approach to AI-assisted work with the wider developer community.

He has published 23 Claude Code skills under the name gstack on GitHub, providing a collection of tools and workflows associated with his use of AI agents.

The project reflects an important development in the AI ecosystem: users are increasingly looking beyond basic chatbot conversations and experimenting with specialised instructions and workflows that allow AI systems to perform particular tasks.

Claude Code is an AI-powered coding tool designed to assist developers with software engineering activities. Skills and reusable workflows can help structure how such tools approach different assignments.

Instead of repeatedly explaining the same instructions, developers can use predefined workflows to make AI assistance more consistent across tasks.

Tan’s decision to publish gstack also provides an opportunity for other developers to examine and experiment with the methods he uses.

This reflects a broader shift towards sharing AI workflows rather than simply sharing individual prompts. As AI agents become more capable, reusable task structures could become an increasingly important part of how individuals and organisations use the technology.

Why Garry Tan Says AI Agents Are Still Too Expensive for Everyone

Despite his extensive use of AI agents, Tan does not believe that this level of personal AI assistance is affordable for everyone yet.

He has acknowledged that the cost of running AI agents remains high and will need to decrease before such systems can become mainstream.

His own annual expenditure of up to $200,000 highlights the gap between what advanced users can achieve through intensive AI usage and what ordinary consumers may reasonably afford.

The cost challenge is particularly important because personalised agents often need to process information repeatedly, access different sources and perform multiple operations to complete a task.

As these systems become more integrated into everyday activities, their operating expenses could become an important consideration for both individual users and businesses.

Why Token Costs Matter for the Future of AI

The economics of AI services depend partly on how much computing power is required to process requests and operate models.

When users rely on AI for occasional questions, the cost may be relatively manageable. But when agents continuously examine documents, retrieve information, generate responses and execute multi-step workflows, usage can increase significantly.

Several factors influence the overall cost of an AI-agent setup:

• Frequency of use: Agents that operate throughout the day can consume more tokens than tools used occasionally.
• Amount of information processed: Large documents, extensive email archives and long conversation histories can increase processing requirements.
• Complexity of tasks: Multi-step assignments may require several model interactions before completion.
• Model selection: Different AI models and service providers may charge different rates for processing and generating information.
• Continuous automation: Systems that repeatedly monitor information or perform background tasks can generate additional usage.

For AI agents to become accessible to a much wider audience, providers will need to make their services more affordable while maintaining reliability and performance.

AI Agents and Personal Privacy: The Biggest Concern Behind Tan’s Setup

Tan’s AI arrangement also highlights one of the most important questions surrounding personalised artificial intelligence: how much personal information should an AI system be allowed to access?

Giving an agent access to emails, cloud storage and personal notes can make it more useful. However, granting access to therapy notes and decades of private records introduces additional privacy considerations.

Such documents may contain sensitive information about relationships, personal experiences, professional disagreements and other aspects of an individual’s life.

The more information an AI agent can access, the greater the importance of understanding how that information is stored, processed and protected.

Convenience Versus Control

Personalised AI assistants create a trade-off between convenience and control.

Broad access allows an agent to understand more about its user and potentially provide better contextual assistance. Restricted access, meanwhile, can reduce the amount of personal information available to the system but may also limit its usefulness.

Users considering similar systems need to understand the permissions they grant and whether an AI service can retain, process or share connected information under its applicable policies.

Another important consideration is whether the agent can take actions independently or only provide recommendations. Reading an email archive is different from having permission to send messages, modify documents or make decisions on someone’s behalf.

For personalised AI to become widely trusted, users will need meaningful control over both the information an agent can access and the actions it can perform.

Garry Tan Wants People to Have AI Systems They Can Control

Tan believes that people should eventually have AI systems they can control and customise instead of depending entirely on technology companies to manage their digital assistants.

This vision reflects an important debate about the future ownership of personal AI.

Many AI assistants currently operate through services controlled by technology companies. Users interact with the models and applications provided by these businesses, while the companies determine important aspects of the infrastructure, data handling and available features.

Tan’s preference points towards a model in which individuals have greater control over how their AI assistants operate and what information they can access.

Such control could become particularly important as AI agents gain access to increasingly sensitive personal and professional information.

However, building a genuinely user-controlled AI ecosystem will require more than making models available. It will also involve secure data management, clear permission systems, reliable computing infrastructure and tools that allow users to understand what their agents are doing.

Cisco Gives a 27-Year-Old Employee Unlimited AI Tokens for a Major Coding Project

While Garry Tan’s experience demonstrates how expensive personal AI-agent usage can become, another example from Cisco shows how companies are experimenting with large-scale AI adoption inside their engineering teams.

Cisco has reportedly given a 27-year-old employee unlimited AI tokens for a project aimed at rewriting a substantial portion of the company’s legacy software code.

The employee, Padmanabhan Krishnamurthy, popularly known as Paddy, has been given the freedom to use as many AI tokens as necessary for the assignment.

Cisco president and chief product officer Jeetu Patel discussed the initiative in an interview with Fortune, explaining that the employee had been given considerable flexibility to use AI tools for the engineering project.

Patel said: “I’ve given him unlimited tokens because he’s going to refactor code. So I’m like, you can go, you can go nuts.”

The decision illustrates a different approach to AI spending. Instead of limiting an employee through a fixed token budget, Cisco is allowing him to use the available technology as extensively as required to demonstrate what AI-assisted engineering can achieve.

Who Is Paddy and Why Did Cisco Give Him Unlimited AI Access?

Padmanabhan Krishnamurthy joined Cisco in 2025 after the company acquired EzDubs, a real-time translation startup he co-founded in 2023.

According to Patel, he came across Paddy working in a role that did not fully utilise his abilities.

During a conversation, Paddy proposed that he could rewrite Cisco’s legacy code with a team consisting of just two people.

The proposal attracted Patel’s attention because of the scale of the challenge and the possibility of using AI to change how large software engineering projects are approached.

Patel subsequently gave him four weeks to demonstrate the proposed approach.

Giving an employee unlimited tokens for such a project represents a willingness to prioritise experimentation and results over conventional restrictions on AI usage.

It also suggests that Cisco sees AI-assisted software development as a potential way to improve engineering productivity, particularly when working with large and complex legacy systems.

Cisco’s Target: Refactor 98 Per Cent of Legacy Code Within 24 Months

The larger objective of the project is ambitious. Cisco aims to refactor 98 per cent of its legacy code within 24 months.

Legacy code refers to software that has been developed and maintained over time but may rely on older programming approaches, structures or technologies.

Large technology companies often maintain extensive software systems that have evolved through years of updates and modifications. Improving such systems can require considerable engineering effort because changes must account for existing functionality and dependencies.

Code refactoring involves restructuring existing software to improve its organisation, maintainability or efficiency without necessarily changing its intended behaviour.

For Cisco, the proposed target represents an opportunity to examine whether AI tools can accelerate this work at a scale that would be difficult to achieve through conventional engineering methods alone.

However, rewriting or refactoring a large software codebase also requires careful testing and validation. AI-generated code must still be reviewed to ensure that it works correctly and does not introduce security vulnerabilities or unexpected problems.

The success of the initiative will therefore depend not only on how much code AI can generate but also on whether the resulting software meets the company’s engineering and reliability requirements.

What Garry Tan’s AI Spending and Cisco’s Experiment Reveal About the Future of Work

The experiences of Garry Tan and Cisco demonstrate two different but connected directions in the adoption of artificial intelligence.

Tan is using AI agents as highly personalised assistants with access to extensive personal and professional information. Cisco, meanwhile, is experimenting with AI as a tool for accelerating complex engineering projects.

Both examples suggest that AI is moving beyond occasional productivity assistance towards more deeply integrated workflows.

For individuals, the potential benefit lies in reducing the effort required to manage information, organise responsibilities and prepare for important decisions.

For businesses, AI could help accelerate software development, improve internal information retrieval and allow smaller teams to tackle ambitious projects.

However, these possibilities come with practical limitations. High token costs, data privacy concerns, reliability issues and the need for human oversight remain important challenges.

Tan’s spending also demonstrates that the most advanced AI experiences are not necessarily affordable for ordinary users. Until operating costs decrease, access to highly personalised agents may remain concentrated among people and organisations willing to spend heavily on the technology.

Cisco’s experiment, meanwhile, will be judged by whether its AI-assisted approach can deliver measurable improvements in software engineering while maintaining quality and reliability.

The broader shift is clear: artificial intelligence is increasingly being treated as an active participant in everyday work rather than simply a tool for answering questions. Whether personal AI agents eventually become affordable and trustworthy enough for mainstream adoption will depend on how effectively technology companies address the challenges of cost, privacy and user control.