For the past seven days, I've had an AI agent take over my email inbox. It can read all my messages, create email drafts, and even send emails on my behalf. The agent has handled more than 60 emails I received during this period.

Initially, the AI agent had no prior knowledge about me and my typical behavior, which caused some issues. However, through a few iterations, we were able to significantly improve its performance. The agent read through hundreds of my past emails to extract facts and knowledge about me, such as my Twitter, Discord, and location. It also learned my voice, tone, and logic for responding to messages based on these past examples. This allowed the agent to create a digital version of myself, behaving almost identically to how I would.

The agent also developed a sophisticated behavior pattern, where it first categorizes each new email and then takes different actions based on the type of category. For example, for emails related to partnerships, the agent will provide a summary about the prospect, their company, and check my calendar to coordinate a meeting time that works for both of us. If the case is beyond the agent's capabilities, it will escalate the email to me.

In this section, I will share the key things I learned from having a real autonomous agent manage a core aspect of my life, as well as how you can build your own super AI assistant.

To create a second knowledge base of extracted knowledge and facts about myself, I took the following steps:

1. I exported all my past sent emails from Gmail and converted the Mbox file into a CSV file with the email text and my responses.

2. I used a large language model (GPT) to parse each email response and extract key facts and knowledge about myself, such as my Discord link, where I live, etc. This was done by breaking the email responses into smaller chunks, passing them through a prompt to extract FAQs, and then combining the results.

3. I saved the extracted FAQs into a new CSV file, which became the second knowledge base for my AI assistant.

Now, whenever my AI assistant receives a new email, it can reference both the email response history as well as the extracted facts to provide a more personalized and accurate response, mimicking how I would normally respond.

This two-pronged knowledge base approach allows the AI assistant to have a deeper understanding of my identity, preferences, and typical communication style, making it a more effective and autonomous agent.

The whole reason I started building this AI assistant is because my email inbox was getting out of control. I was getting a lot more email inflow than I could handle, and sometimes I would see the email and know the response, but I wanted to make sure the email was properly written. Normally, I would use ChatGPT, give some context about myself, the original email, and ask it to generate a response, but this process was quite time-consuming, and when I was only with my phone, I couldn't really do this because copy-pasting things between different apps was not that straightforward.

This got me thinking, what if I can have GPT to just create a draft response for every single email every morning, whether I check my email on the computer or on my phone? A draft response is already there, and I just need to review and click Send. That's exactly what I did. For the first version, I didn't even write any code; I literally used Zapier to build this workflow where I would trigger GPT whenever I receive a new email with a specific prompt, and once GPT returns the result, it will create a draft reply. The reply process is a little bit complicated, as I need to search for the email with the email address and subject first, which gives me the thread ID that I can use to create a draft to reply and pass on the body response. I set this up in 5 minutes, but the result was amazing - every single new email had a draft created already, and I could directly click on that, review the results, make some edits if I needed, and click Send. It was accessible anywhere, even when I was on my phone.

This was my first learning - by bringing AI into a user's existing flow and making it contextual, it can be a magical experience. However, the problem was very obvious because this was simply one GPT prompt, and it had no context whatsoever about who I was. It definitely caused a lot of problems, and 50% of the time, the response quality was so bad that I couldn't really use it, and it would also hallucinate and make the wrong decisions for me. To make this AI assistant useful, I had to teach it some facts about myself and the domain I operate in, so it can answer the questions like I do.

This came the second iteration, where I wanted to digitize and create a knowledge base about myself so the AI assistant can mimic my behavior. The approach I decided to take was creating a knowledge base for my past emails - whether I can extract the facts and knowledge about myself just from past emails and also create a database so whenever the AI assistant receives a new email, it can try to look up what was the last time I received a similar type of email and how did I respond before.

I first explored all the past emails that I've ever sent and then extracted clean data into a CSV file that has two columns: one original message that people sent me, and the other is how I responded before. Then, I took the original reply and turned it into an extracted facts and FAQ in JSON format. I then used this two-main knowledge source to do a vector search whenever I receive a new email so that this AI assistant can actually have a reference to learn from.

The result was amazing. I was able to create a sophisticated email categorization and response system where the AI agent can now categorize each email into different types, such as consulting requests, collaboration opportunities, or general inquiries, and take different actions based on the category. For example, if it's a consulting request, it will first check if the email has provided the necessary information like the problem they're trying to solve and their budget, and if not, it will generate a response to collect that information before escalating it to me. If it's a collaboration or sponsorship email, it will research the company and the opportunity first before forwarding it to me with a summary.

This system has been a game-changer for me, allowing me to efficiently manage my email inbox and free up my time to focus on more important tasks. By leveraging AI and building a robust knowledge base, I've been able to create a personalized and intelligent email assistant that can handle a significant portion of my email workload.

The key takeaways from this experience of having an AI agent manage my email inbox are:

1. Integrating AI into existing workflows can provide significant productivity gains, even with simple implementations.
2. Building a comprehensive knowledge base about oneself, by extracting information from past emails, allows the AI agent to better mimic one's behavior and respond more accurately.
3. Developing a modular, customizable system of specialized tools enables the AI agent to handle a wide range of email scenarios with sophisticated decision-making capabilities.
4. Continuous updates and refinements to the knowledge base and tool set are crucial to maintain the agent's effectiveness over time.

Overall, this project demonstrates the potential for AI to automate and streamline personal administrative tasks, freeing up time and mental energy for higher-value activities. With the right approach, an AI assistant can become a seamless extension of one's own workflow and communication style.