Vitali Nikulenka · November 2026

One person and a team of AI agents

I have been in IT for 38 years. In 2026 I stopped using AI as a chat window and built a standing team of agents. The team gathers information, drafts documents, writes and tests code, and keeps track of deadlines. I keep the task-setting, the checking and the decisions.

Mail Code Meetings Docs Deadlines Data Me
20projects in progress per month. In April there were four
69%of my changes to projects are made together with an AI agent
115project dossiers kept up to date by agents
110open commitments tracked by an agent in one summary drawn from 13 sources
How I work

Who does what

Every task goes through the same loop. Two of the five steps are mine, three belong to the agents.

my partthe agents' part
Me

I set the task and the rules

Agent

Does the work

Agent

Checks itself: tests, cross-checking against primary sources

Me

I review the result and decide

Agent

Records the outcome in the knowledge base

Each loop leaves a trace in the knowledge base, so the next one goes faster.

My rule

I hand to AI what is easy to check. I do it myself when checking costs more than the work itself, or takes expertise the checker doesn't have.

How my work with AI has changed

1
ToolI ask a question and get an answer
2
AssistantI hand over part of the work
3
AgentAI carries out a chain of actions on its own
4
Part of the process I am hereAI is built into how the company works
5
Personal layerAI is always there and takes over the routine
Personal automation

How it is set up

The system has three parts. Agents on an always-on computer in the office work in the background and message me on Telegram. Claude Code on my laptop helps with documents, analysis and code. The “second brain” is a folder with a dossier on every project: any agent picks up the context from it within a minute and doesn't ask me about what is already known.

Inputs

Telegram
Email, 4 mailboxes
Calendar
Tasks and reminders
Meeting recordings
CRM
Code repositories
Files on disk

What runs

Hermes agentsAn office Mac mini, running around the clock. 9 profiles: mine and my colleagues'
Claude CodeOn my laptop, we work side by side
Second brain115 dossiers, refreshed every morning

What I get

A “what do I need to do today” summary
Meeting minutes with action items
Tender digest
Dashboards for executives
Presentations
Drafts of emails and proposals
Working code

My day with the agents

An agent has gone through new public procurement notices and sent me the relevant tenders

An agent has gone round every project and updated the dossiers

I ask: “What do I need to do today?” An agent pulls together my commitments from email, Telegram, the calendar and the CRM

Recording → transcript → minutes → action items for participants

Documents, analysis, presentations and code together with Claude Code

Dashboards for executives update themselves

Division of labour

AIMe
Gathers and consolidates informationSet the task
Transcribes meetingsSet the rules and acceptance criteria
Writes the first draft of a documentReview
Writes and tests codeMake the decisions
Finds sources and checks factsTalk to people
Keeps track of deadlinesAnswer for the result

What I don't hand over unsupervised

  • Emails and messages in my name go out only at my direct request
  • In services built for clients, personal data passes through an anonymiser before it reaches a cloud model
  • Meeting recordings are transcribed on our own hardware
  • My colleagues' agents have no access to the terminal or to other people's files
Result

What it gives me

The number of projects I run at the same time grew fivefold in six months: 4 in April, 20 in October.

Projects in progress per month, 2026

FebMarAprMayJunJulAugSepOctNov

A project counts as in progress if my changes appeared in it during the month. Tracking starts in February 2026, when I moved my work into repositories. * November covers the first 20 days. Over the same months the number of dossiers in the “second brain” grew from 20 to 115.

1 dayFirst version of the meeting-recording system, 129 automated tests
1 dayStage one of a portal for access to AI models, 175 tests
3 daysTesting 28 trading rules in 112 variants on historical data
Projects

What is built on it

ParkLab Technologies

AI Digest

Media and social-media monitoring for executives. In use at the Secretariat of the Supervisory Board of the Hi-Tech Park (Belarus). 14 banks and organisations are on their way to a 30-day trial.

In production
My role: commercial director, I take the product to clients
ParkLab Technologies

Kontur 01

An assistant portal for the executives of the Hi-Tech Park Secretariat: chat with data protection, a legal assistant for legislation, a procurement radar, meeting recording.

In production since October 2026
My role: architecture and development together with AI agents
Own product

Meeting recording

Recording → transcript → minutes → action items. The data never leaves our own hardware. Runs inside the Kontur 01 portal and on an office server.

Two installations
My role: designed and built it
CTT · EDI provider, 9,235 clients

AI at CTT

A document-flow dashboard for executives, AI summaries of deals for 346 companies in the CRM, a procurement radar, a presentation service for sales managers.

Live services and pilots
My role: director of innovation, conceived and built it
UNUM

Trading bot

Strategy research validated on data the model has never seen. Paper trading since 26 October. The decision on a real-money pilot is due on 8 December 2026.

Paper trading
My role: validation method and technical architecture
Hermes

Personal agents for colleagues

9 profiles on Telegram: executives and managers, plus the first paying user. The agent transcribes voice, works with documents and reminds about deadlines.

Running
My role: deployed and maintain it