As of today, I’m the last human in my team.

I replaced all of my teammates with AI agents.

So, first thing, what are AI agents? Is an agent something like this:

Probably not. That’s just a n8n workflow: a glorified reproducible, shareable software

function. Just a bunch of if/else conditions.

For me, an agent is an autonomous process that is capable of learning and deciding what steps to take. So, an agent is deciding and taking action, while learning in the process.

For the longest time I though of agents as a job to be done. Want to follow up with leads? —> Let’s build an agent for that

Want to analyze your website visitors? —> Let’s build an agent for that

Then, I started thinking of agent as roles to be done.

That’s why any of my agents can perform multiple actions at once.

Not only that, I treat them like humans! Each of my agents has:

  • Their own email address
  • Phone number
  • Messaging profiles
  • Memory
  • Specialized tools
  • Decision-making power

I design my agents to be proactive. They multitask.

They get things done faster than I ever could.

—

How does it work?

The infrastructure is ever-changing, but this is a good picture of the status in May 2025.

This infrastructure can understand context from multiple sources, learn from interactions, and take actions across various platforms and services.

Core Processing Flow: The system starts with a user question that gets processed by a central Brain API. This API coordinates the entire workflow through several sequential stages: Learnings → History → Scoring tool → Retrieval tools → Action tools → Answer.

Input Channels: I can communicate with my agents choosing any of the input interfaces: our own web app, Telegram, Mattermost (our Slack-alternative company chat) , VAPI (for voice calls), email, and WhatsApp, allowing to interact through my preferred medium. Most of the input interfaces connect to an auth system that handles biometrics and code-based authentication, ensuring secure access.

Learnings and history:

The first thing we do when we receive a new conversation is to retrieve the chat history for that agent, and the learnings that agent already stored.

The scoring tool checks if we can reply to the question just based on history and learnings, or if need retrieval tools to collect the information needed to formulate a reply.

Retrieval tools: RAG tools are both sync and async.

  • Sync tool might include the weather tool, the browser tool, the scraping tool, the calendar tool, etc..
  • Async tools might include data that is stored on a recurring basis from services like Notion, Oura, Tasks, Rescuetime, Playtomic, FatSecret, Meals.chat, Strava, and more

I use both relational database and vector database (embeddings systems) for knowledge management. Each tool might use the MCP standard or not, and I might use different LLMs for different tools, all monitored with LiteLLM for observability.

Action Tools: If there is a need for an action to be taken by the agent (again, an agent must have the ability to do so!), we have action capabilities including a LinkedIn tool, a n8n tool (automation), a GitHub tool, an email tool, a calendar tool, and so on. This allow the agents to not just retrieve information but also perform actions.

—

Meet the Team

Curious to meet my coworkers? Let me introduce you to a few of them:

  • Carla, my AI Chief of Staff: She runs my schedule, manages tasks, creates SOPs, follows up on leads, and even checks in on me.
  • Ella, AI Financial Controller: Knows the finances of each product and gives proactive reports.
  • Gerry, AI Head of Product: Tracks product health, analyzes competitors, and generates insights.
  • Lee & Sarah, AI Sales Reps: They handle demos, and optimize sales conversations—24/7.
  • Herbie, AI Developer: Fixes bugs, writes tests, builds new workflows, and understands my coding style.
  • Chick, AI Security Lead: Audits code, flags vulnerabilities, and watches user paths.
  • Keith, AI Sysadmin: Keeps the servers breathing and stable.
  • Toshiko and Bill, AI Content & Marketing Heads: Write and distribute content, send campaigns, and adapt strategies based on user profiles.

Are they better than a real team? I don’t know. I never had a real one.

But it’s more people than I ever had.

I’ll go in more detail about each of them in separate blog posts.

Thanks for reading,

Mike.