MICHAŁ PAWLUKIEWICZ
What I built

Fourteen builds. Each one ran on my own setup first.

Since May 2026 I have been building with AI on my own work, my own infrastructure, my phone and my car. Every build is listed below: what it is, how it runs step by step, what the measurements showed, what else it can do and which principle it implements. The numbers were counted by the agents that run each build.

  1. System for working with agents
  2. Agent in a messaging app
  3. Phone controlled by an agent
  4. AI in the car
  5. Local model with control
  6. Infrastructure run with AI
  7. Audit-grade backups
  8. Shared dashboard
  9. Simulator game
  10. AI journal
  11. Overnight app tests
  12. Company workflow: 8 agents
  13. Voice assistant
  14. Method and learning
Agent team

24 agents, each with its own area

These are not 24 chat windows. This is a team with divided roles, where nobody changes someone else's work and everyone reads the same principles and the same failure log.

Company workflow · 8

  • dispatcher for mail and decisions
  • hardware selection for the quote
  • deployment pricing
  • decisions from the phone
  • live workflow dashboard
  • workflow site and map
  • demo
  • data from the ERP system

Agent server · 7

  • machine log and updates
  • host: ports and load
  • guard for the messaging-app agent
  • shared dashboard
  • phone controlled from the computer
  • devices and hardware analysis
  • overnight app tests

Main computer · 9

  • overall supervisor
  • infrastructure and network
  • AI in the car
  • local model
  • data from the ERP system
  • method and education
  • project housekeeping
  • this site
  • backups
01 · foundation

System for working with a team of agents

The rules, memory and failure log that 24 parallel agents work by. Everything below rests on it.

Principles: 01, 03, 04

  1. I set the goal and make the decisions. Code is written only after my “build”.
  2. The agent shows the plan, estimates time and cost, and hands larger tasks to subagents.
  3. It checks the result from the end, the way I will use it. Otherwise it writes “UNVERIFIED” and what is missing.
  4. A failure that matters for the project goes into the shared log: what failed, what to do next time, what it cost.
  5. Every area has one owner. An agent does not change someone else's work. It reports it.
182failures written down in the archive
11mechanisms I reduced them to
24agents under one set of principles

Can: onboard a new agent in a few minutes. It reads the principles, the project memory and the log, and then knows what belongs to it and what it must not change.

02 · access

Agent in a messaging app

I write from my phone to a separate number and an agent running on my server replies. I have it with me everywhere, without opening a computer.

Principles: 02, 06

  1. The message arrives at a dedicated number, separate from my private one.
  2. A guard script reads it and starts the agent with a restricted set of tools.
  3. The agent has read access to mail, the drive and selected folders. It changes the calendar only after my “yes”.
  4. Files go both ways through a filter: no keys, password databases, macros or executables.
  5. Simple commands are handled by the script itself, with no AI and no cost.
14 daysof continuous operation: 0 lost connections, 0 restarts
15 / 15questions to the agent answered; 16 commands with no AI involved
7×cheaper after a session reset (193k → approx. 28k tokens per call)

Can: pass an instruction to a specific agent at the computer, take a photo and describe it, send back a file from the drive, show a snapshot of the whole environment.

03 · hands

Phone controlled by an agent

A separate phone that the agent controls like a second computer: it sees the screen, touches it, takes screenshots. It is used to test apps on a real device and as a sensor in the field.

Principles: 02, 03

  1. The agent connects to the phone wirelessly and selects it by serial number, so devices cannot be mixed up.
  2. It reads the screen, performs one touch, takes a screenshot. Only then the next step.
  3. When a command does not return, it reads the screen first. Usually it is a system permission prompt.
  4. Without my “yes” it installs nothing, makes no calls, sends no SMS and changes no settings.
28,627route points collected in 17 days
1,448 / 1,449successful connectivity checks, every 15 minutes
approx. 1 sapp response after the agent touches the screen
40 screenshotsin two app tests on the phone

Can: walk through any mobile app like a user and deliver a report with a screenshot after every step, without an emulator. Take a photo with the camera on a command from the messaging app, show its position, report movement every five minutes.

04 · field

AI in the car

A phone left in the car reports “moved off” and “parked” by itself, shows its position on request and wakes up on command. A layer alongside the factory systems. Drivetrain, brakes and steering are outside AI.

Principles: 03, 05, 07

  1. A local agent on the phone judges movement and standstill from GPS points.
  2. Data packets go over the private network to a receiver on my server.
  3. The server turns events into reports with a place name.
  4. It detects ignition from the connection to the car. On a long standstill it puts the link to sleep and wakes up on command.
  5. A separate technical channel carries alarms and buttons, with no location.
1,187.7 kmin 79 drives, 20 days with drives (27 Oct–15 Nov)
172“moved off” and “parked” reports
3,121 / 3,122points delivered on a 320 km route (31 Oct)
10.2 → 1.44 %/hbattery drain when parked, after fixes
36 sfrom command to execution out of sleep

Can: take a photo of the interior on request, signal with light and sound, show the day's route, resume work after a restart, unattended, in a minute and a half.

05 · data

Local model with a control layer

My own language model on my machine, with no graphics card, called with a single command. The core is the control layer: task scope, a sandbox and result checking by a second model.

Principles: 02, 03, 06

  1. The question goes to a gateway that keeps history, logs and picks the model for the task.
  2. A task has a scope and a measurable “done” criterion. The loop is limited to five turns.
  3. The model's code runs in a single-use container: no network, read-only, memory and time limits.
  4. The result is judged by a second model, not the author. I do not trust “done” declarations.
  5. Mail anonymiser: the data stays local, the content goes out without it, a human approves the draft.
288.6 → 62.1 sthe same series of questions after choosing the variant
6 / 6sandbox tests: loop interrupted, network cut off, memory limit held
29 / 30personal data items detected; amounts preserved in 100%
20 attacks0 leaks in replies, 3 through the tool channel; 0 after access was cut off

Can: read and create Word and PDF documents in the sandbox, talk through a browser window, serve three models (general, Polish, reasoning) behind one gateway. The data does not leave the building.

06 · infrastructure

My own infrastructure run with AI

Eight devices in a private network: computers, servers, virtual machines, phones. Agents start services, change rules and watch the backups. I approve.

Principles: 01, 02, 05

  1. Every connection in the network must be explicitly allowed. One agent changes the rules, after my “go”.
  2. Services listen locally only. They go out exclusively through the private network.
  3. Every change has a “what and how to undo” entry and goes into a private configuration repository.
  4. I type passwords myself, in a window the agent opens for me. The agent does not see them.
  5. Agents working around the clock got a separate server, so they do not load my work computer.
313recorded environment changes, each with a rollback path
173changes in the configuration repository
version 20of the private network rules
55 minto move the agents to a new server
24 vCPU / 96 GBmachine for my own services

Can: start a new service from scratch in the private network in one session, with a log entry and a ready rollback path.

07 · continuity

Backups with a restore test and audit-grade documentation

The backup server covers ten machines and one report arrives in the morning. For the upgrade, the agent prepared full validation documentation in the style of regulated industries, in Polish and in English.

Principles: 02, 03

  1. Backup jobs run at night, separately for servers, virtual machines and computers.
  2. In the morning a script assembles one combined report, and one line goes to the shared dashboard.
  3. The check is a restore and a checksum comparison, not a green status.
  4. The agent is the author and tester of the documents. It never signs them.
10 machinesin six backup jobs
9 / 9files restored with matching checksums
2 / 2full virtual machine restores passed
15 + 2documents and procedures in two languages, 27 deviations described
88.96 s → 0.0019 sclock drift detected and fixed

Can: start a whole virtual machine straight from a backup and document it so that it passes a quality inspection.

08 · visibility

Shared dashboard for the whole environment

One page that shows everything: messages, code changes, machine state, AI chats, the car and the backups. A mirror and a log. Decisions are not made on its basis.

Principles: 03, 04

  1. After an important change the agent adds one trace line, with no addresses and no secrets.
  2. Collectors gather the state of the network, machines and routes, read-only.
  3. The phone gets a static snapshot in a single file.
  4. Alerts arrive batched, at most once every four hours, with quiet hours at night.
2.5 hto build the second version
771 / 772tests pass
105 → 35 MBmemory after the fix, flat across 480 requests

Can: show on the phone the state of 24 agents, eight devices, the car and the backups, about 20 seconds after the command.

09 · code

Command simulator game

A game in C#, one executable file. It lets you feel in practice the thesis that an AI arbitration layer gives tempo while the human keeps control and the right of veto.

Principles: 01, 03

  1. The change plan goes into the change file before my “build”.
  2. The agent writes code in one of the two editions of the game and builds with a native compiler.
  3. A headless bot plays the missions and checks the balance with one script.
  4. Release: a change in the repository, a package with version and date.
~60klines of C# per edition of the game
4 daysfrom the first session with the agent to two releases
4,174agent actions in one day, on release day. The record

Can: test itself. The bot plays all the missions and says whether the balance holds before a human sees it.

10 · memory

AI journal, day by day

A complete chronicle of my work with AI since the first conversation: every day, every conversation as a story, a point and a comment. It builds up with no effort from me.

Principles: 04, 07

  1. Once a month I order an export of the conversations.
  2. Scripts split the raw material into days. The second stream is the records of sessions with agents.
  3. AI subagents write the days in parallel, to one set of guidelines.
  4. The agent assembles weeks and volumes, scans for secrets and marks private threads.
5 volumes24 weeks, 137 days of work described
17,560messages in 435 conversations, plus 163 sessions with the agent
79 agentsproduced the whole, approx. 21.1m tokens
20 agentsone month in under 15 minutes

Can: answer what I was doing on any day since 29 May 2026 and what conclusion I reached then.

11 · quality

Overnight app tests by agents

Agents walk through someone else's app like a user, in a browser and on a real phone, with a screenshot after every step. The report is waiting in the morning. My business partner writes the code, I run the direction and the tests.

Principles: 01, 03

  1. A new version arrives with the author's note.
  2. Several agents walk through scenarios on their own copy of the app, with no changes to the author's version.
  3. The output is a report with screenshots and a purely descriptive note with numbered remarks.
  4. I read, decide and send the note myself. The author's repository is read-only for the agents.
15 testsincluding 3 overnight, 56 scenes described
324screenshots in the report from one night
35 + 24remarks and proposals in the first report
77points in the code review

Can: take an app in the evening and deliver in the morning a report of hundreds of screenshots, with a numbered list of remarks.

12 · company

Company workflow: eight agents from email to quote

The same approach applied to the work of a company. Four agents take an enquiry from the email to the sent quote, four more provide a live dashboard, a workflow map, a demo and data from the ERP system. A human approves at three points.

Principles: 02, 03

  1. The dispatcher receives the email, identifies the case and distributes the work.
  2. The quote agent selects hardware from a catalogue of 2,667 items and proposes substitutes.
  3. The deployment agent prices the configuration and draws the diagram.
  4. A human approves, corrects or rejects. Three decision stages.
  5. The dispatcher sends the quote and leaves a trace of every decision.
17 and 29 minfrom email to quote on two real enquiries measured
approx. 3 minof human work in the whole run
46emails handled, 3 quotes sent
362automatic checks of the workflow in the sandbox
36 snew version after “give a 5% discount”
25 / 25report figures reconciled with the ledger

Can: take a decision from the phone (approve, correct, reject), recognise a case of an unknown type and hand it to a human, assemble a report for the management board straight from the ERP system.

13 · voice

My own voice assistant on a PLN 70 module

A 24 × 24 mm module with a microphone and a speaker that listens for one word and lets me talk to the assistant by voice, at home and in the car. It is the ear and the mouth. The thinking is done by my server, not someone else's cloud.

Principles: 02, 05, 07

  1. The chip recognises the wake word by itself. Until that moment nothing leaves the device.
  2. After the wake-up the voice goes over the network to my server.
  3. The server turns speech into text, the model replies.
  4. The reply comes back as a Polish voice from the module's speaker.
1 dayfrom delivery to working open-source firmware
2 × 8 MBbackup of the factory firmware, two reads with an identical checksum
28.6 sfull restore of the factory state, verified
PLN 69.90list price of the hardware

Can: talk in Polish after the wake word, remember up to ten networks (home and car), trigger actions on the phone. No voice goes to the manufacturer's cloud.

14 · people

Method and learning programme for working with AI

Everything above can be repeated in a company if the people know how to work with AI. So I wrote down the method and put together a learning programme in which education is the first stage of the work, not a training course on the side.

Principles: 01, 03, 07

  1. First the alignment: “how do you understand this?”. Only then action.
  2. An unknown is marked explicitly instead of guessed.
  3. Every decision leaves documentation.
  4. The proof of skill is the trace of the work, not a theory test.
  5. Memory and data belong to the company, not to the tool.
8 stepsof the learning programme, based on the AI Fluency model
13 chaptersof the written method
35cases in the test of the workflow between chat and repository

Can: build a company a data gateway in front of AI, a test range with mock systems for regulated environments, a sandbox with virtual customers, the “company answers questions about itself” test and the “no-AI mode” test.

Flow

Three paths from build twelve, step by step

Pick a case. Green is done by AI, gold is decided by a human, red is out of the model's reach.

Failures

Four failures that became rules

Anyone who shows only successes is not measuring. These four changed the way I work.

Overnight work stopped for three hours

The automation that was meant to resume the agent's work stopped at a permission prompt. Nobody could grant it, because it was five in the morning.

Rule: automation must not depend on a question to a human who is not there. Permissions are recorded in advance for specific tools.

Four times “success”, zero emails sent

A scheduled task ended with the status “successful” four days in a row. Not one email went out.

Rule: check the effect, not the status.

The dashboard was down for two and a half days and nobody noticed

A closed console stopped the service that was meant to show the state of everything else.

Rule: the dashboard is a log, never a source of truth. Agents keep working when it is unavailable.

An alarm that had no right to fire

The “no backup report” rule sent a warning although it had no data source connected. Fixed the same day.

Rule: every new alarm goes through a dry run on real data before it starts sending notifications.

Where this comes from

I run agents the way I ran the group's infrastructure.