Last Updated: March 5, 2026
At TechChill 2025, I delivered a keynote about where AI is heading for entrepreneurs — from managing your inbox with simple automations to multi-agent systems that replace five months of full-time research work in two hours. What I could not share publicly at the time was that I had already sold a multi-agent system to a Tier 1 US financial institution. Two years under NDA, learning everything about how these systems work at scale. That project is behind me now. But the lessons I took from it shaped everything I built after — including GoBot, the AI Second Brain that now runs my business. This keynote captures the moment I first shared this vision publicly, and everything I predicted has since come true — faster than expected.
From a Bank Deal to Building in Public
At the end of 2023, I delivered a keynote at the European Investor Summit. In that audience was the head of innovation at a Tier 1 US financial institution. That talk led to a deal. I spent roughly two years involved with a multi-agent system deployment for them — under NDA, unable to share a single detail publicly.
That project is finished. I am no longer involved. But what I learned during those two years fundamentally changed how I think about AI architecture. The patterns for preventing hallucination. The reasoning techniques that actually work under pressure. The safeguards against meta-prompting and infinite loops. How to structure memory so agents do not pollute each other’s context.
I took those lessons and applied them to GoBot — the AI Second Brain I built for myself using Claude Code. It is a completely different system, built from scratch, designed for entrepreneurs and professionals instead of enterprise banking. But the architectural principles are the same.
The TechChill keynote was the first time I spoke about multi-agent systems publicly. Here is what I shared, why it matters, and how it all aged like fine wine.
The Core Idea: Thought to Action
We spend our days doing two things on our devices. We create and transform data — writing, editing images, producing videos. And we move data around — sending emails, posting content, dragging files between folders.
AI already does both of these things better than human hands on a keyboard. And it gets better every single day.
The trajectory is clear. We went from typing with fingers, to voice commands, to AI assistants that handle multi-step tasks from a single message. The next step — already prototyped by companies like Card79 (the team behind Neuralink’s device design) — is thought to action. A device clipped to your ear that lets AI hear your thoughts and act on them.
That is the direction. But we do not have to wait for brain-computer interfaces. What is available today is already transformative.
The Four Levels of AI Adoption
In the keynote, I broke down AI adoption into four levels — designed so anyone can figure out where they are and what to do next. This is the same framework we use inside the Autonomee community to help members avoid overwhelm.
Level 1: Simple Automations (Zero Inbox)
My inbox receives around 2,000 emails per week. Using no-code automations, every email gets classified, labeled, and routed to the right folder automatically. Sales pitches go to a folder I never open. Important messages surface with summaries. Newsletter content gets synthesized into digestible briefs.
This is achievable in an afternoon. No coding required.
Level 2: AI Assistant (One Message, Multiple Actions)
A single message in Telegram triggers multiple actions across multiple tools. “Block time for my panel at 12:40. Set a reminder to buy a gift for Mija. Research how the Lithuanian government uses AI and save it as a note.”
One message. Three different systems. Handled in seconds while I carry on with my day. When I delivered this keynote, deep research as a product category did not even exist yet. Now it is standard — and our setup had it before anyone.
In the community, people set this up in 30 minutes to one hour. From 20-year-old students to experienced professionals who remember when email itself was new.
Level 3: Multi-Agent Systems (Collective Intelligence)
This is where productivity transforms. Not just “how fast can you do something” but “how much can you do.” Multiple specialized AI agents working together, each with expertise in a specific domain, coordinating their outputs like musicians in an orchestra.
As my teammate Jay explained to a client: “Asking a single AI agent to perform multi-stage, multidisciplinary task execution is like expecting a conductor to produce beautiful music by simultaneously playing every instrument themselves. To produce a symphony, you need musicians with expertise in every single instrument.”
I took the lessons from the bank project and applied this exact principle to GoBot. The result is a board meeting system with seven specialized agents — CEO, CFO, CMO, CTO, COO, Research, and Critic — that collaborate to help me make better business decisions. Different system, same architecture philosophy.
Level 4: From 5 Months to 2 Hours
The demo that got the biggest reaction at TechChill. In 2020, I was hired as a project manager to analyze the entire Baltic startup scene. Five months of full-time work. One final report.
Using multi-agent systems, I reproduced a comparable analysis — a 24,000+ page report covering 75 venture capital firms attending TechChill — in two hours.
Not by cutting corners. By using specialized agents that each handled a piece: data collection, analysis, synthesis, insight extraction. Then another agent layer ran compatibility analysis across all the results. The output was not just data — it was actionable intelligence about which investors were actually worth meeting.
Reports like these cost between $5,000 and $200,000 when done by consultants. The multi-agent system produces comparable quality for a fraction of the cost and time.
Want to see multi-agent board meetings in action? I applied the architecture principles from the bank project to build a 7-agent board meeting system in Telegram. Full technical breakdown with setup guide: How My 7 AI Agents Run Board Meetings
Why People Are Overwhelmed (And What Actually Helps)
I surveyed my community about their main obstacle for adopting AI. Two years after ChatGPT launched, the number one reason was not lack of technical knowledge. It was feeling overwhelmed.
Overwhelmed by thousands of options. Overwhelmed by conflicting advice. Overwhelmed by not understanding what is real versus hype.
According to Gartner’s 2026 predictions, 40% of enterprise applications will feature task-specific AI agents by year-end — up from less than 5% in 2025. The technology is moving fast. But the adoption gap is not technical. It is psychological.
That is why we built the community the way we did. You take a quiz. We figure out your level. We tailor what you actually need to learn. No overwhelm. No “you should be doing everything.” Just the next step that makes sense for your business.
The Mija Moment (Why This Matters Beyond Productivity)
The keynote started and ended with my daughter Mija. Three years old at the time.
Her kindergarten teacher asked what her mommy does. She said: “My mom leaves me. And then she goes to the office.”
That broke something in me. But it also clarified something.
What do we actually do in the office all day? We sit at metal machines, clicking buttons, making funny noises. Kids see it clearly because they reason from first principles. They strip away the meaning we adults attach to our work and describe what is actually happening.
The vision behind multi-agent systems is not about productivity metrics. It is about solving bigger problems. If AI handles the file-moving, the email-sorting, the data-transforming — what could we do with that freed-up time and mental energy?
My answer to Mija: “You know how you like making puzzles? We adults love solving puzzles too. We love solving problems. I go to the office because I like to solve bigger problems. But I want that one day, instead of solving small problems like how to save a file, you would have collective intelligence and you could solve bigger problems.”
That is what we are building. Not just AI tools. A world where the mundane is automated so humans can focus on what actually matters.
What Has Changed Since This Keynote
Everything I described on that TechChill stage has accelerated. Here is what is different now:
- GoBot exists. I took the architecture lessons from the bank project and built something entirely new — GoBot, running on Claude Code, self-hosted, with 25+ features including AI board meetings with 7 specialized agents.
- 546+ people are building. The community has grown to 366+ paid members and 546+ free members, all building their own AI Second Brain. Not using my old bank system — building fresh with GoBot.
- The setup got simpler. What took weeks of custom development now takes 30 minutes for the basic relay. The free course walks anyone through it.
- Deep research is mainstream. When I showed that 24,000-page report at TechChill, the audience was stunned. Now every major AI platform offers deep research. We had it before anyone.
- Multi-agent is the standard. Gartner reported a 1,445% surge in multi-agent system inquiries from Q1 2024 to Q2 2025. What was visionary at TechChill is now an industry priority.
Frequently Asked Questions
What is a multi-agent system in simple terms?
A multi-agent system uses multiple specialized AI agents that work together on a task instead of one general-purpose chatbot trying to do everything. Each agent has expertise in a specific area — research, finance, strategy, content — and they coordinate their outputs. Think of it as a team of specialists versus one overworked generalist. The result is higher quality analysis, fewer hallucinations, and the ability to handle multi-stage tasks that single AI systems cannot manage.
Do I need to be technical to build a multi-agent system?
The basic AI assistant setup (Level 2 in the keynote) requires no programming — just copying a repo, setting environment variables, and running a server. The full multi-agent system with board meetings is more involved, but the Autonomee community provides step-by-step courses and live support. Members range from 20-year-old students to seasoned professionals. The community is designed so you do not have to figure it out alone.
How does the thought-to-action concept work today?
While brain-computer interfaces like Card79’s prototype are still in development, the practical version exists now: voice and text commands that trigger multi-step AI workflows. A single Telegram message can block calendar time, set reminders, run research, and save notes — all simultaneously. This is Level 2 in the framework, and most people in the community set it up in under an hour.
Is GoBot the same system you built for the bank?
No. GoBot is a completely different system, built from scratch using Claude Code and designed for entrepreneurs and professionals. The bank project was an enterprise multi-agent deployment under NDA — I am no longer involved with it. What I carried over are the architectural lessons: how to structure agent memory, prevent hallucination across agent boundaries, design reasoning techniques for different domains, and build safeguards against meta-prompting. Those principles informed GoBot’s design, but the code, features, and use case are entirely different.
Watch the Full Keynote
The video above is the complete TechChill keynote with my commentary on how everything has played out since. If you want to see the multi-agent architecture principles I described on stage now fully built and operational, read the technical breakdown: How My 7 AI Agents Run Board Meetings.
And if you are ready to build your own AI Second Brain — start with the free course, or join 366+ professionals inside Autonomee for the full system with board meetings, voice calls, proactive AI, and everything else.
