MICHAŁ PAWLUKIEWICZ
Michał Pawlukiewicz · IT infrastructure and AI

Twenty years of infrastructure under audit. Half a year with AI. 15.9 billion tokens.

I ran group IT in three countries through acquisitions, a company split and an FDA inspection. On 29 May 2026 I wrote my first sentence to an AI. Today I run a team of 24 agents: across a network of eight devices, in a phone, in a car, in code and in a company workflow. Every number on this site was counted by the owner of the data.

  1. first conversation with AI
  2. first session with an agent in code
  3. two releases of a game, approx. 60k lines of C#
  4. an agent answers me in a messaging app
  5. local model in a sandbox, 6 of 6 tests
  6. first 320 km with AI in the car
  7. agents on their own server, moved in 55 minutes
  8. a real customer enquiry: quote in 17 minutes
  9. 1,187 km driven with AI
  10. 24 agents, 67,631 actions executed
Scale

Engineering work, not an experiment

Counted on 19 November 2026 from work records on two machines. The laptop is not included.

15.9 bntokens processed by my agents, including 51.6 million written by the modelstart: May 2026 · counted 7 Oct–19 Nov
67,631actions executed by agents: reads, changes, commands, tests1,537 a day on average
309sessions with an agent and 748 runs of AI subagentsseven sessions a day
24agents working in parallel, each with its own area; eight in the company workflowas of 19 Nov 2026
1,187.7 kmwith AI in the car: 79 trips, 172 reports27 Oct–15 Nov 2026
771 / 772tests of the shared dashboard; first version built in 2.5 hours5–9 Nov 2026
17,560messages with AI written up in five volumes of a journal by 79 agents29 May–29 Oct 2026
182recorded failures of my agents. Every agent reads the log before starting workreduced to 11 mechanisms
Why so fast

An agent is a controlled plant. A control engineer knows what to do with it.

I am a control engineer by training (Automation and Robotics, AGH). When I got my hands on a language model in May 2026, I saw neither magic nor a threat. I saw a plant with a random output that has to be closed in a loop.

Closed-loop control

A system without feedback does not know it is wrong. So the goal is the setpoint, the check in target conditions is the sensor, and the failure log is the feedback. “It works” without a measurement is an open loop.

Stochastic processes

You do not expect repeatability from a random process. You design around the spread. So a second model judges the result, not the author, loops have a turn limit, and a conclusion rests on a series of trials, not on one.

Neural networks

I knew what this was before it became fashionable. Hence work without awe and without fear. And limits built into the construction, because you do not persuade a network with a request.

The theory waited twenty years. Infrastructure supplied the practice. Only the plant was missing.

The difference

Same tool, different class

Almost every company has an AI assistant today. What decides the difference is what has been built around it. Three comparisons that explain it better than numbers.

01 · control

Sat-nav on a phoneAir traffic control tower

Usually

The assistant guides one person along one route. It suggests a turn, answers for nothing and does not know what the others are doing.

Here

Every agent has its own corridor. Take-off needs human clearance, and every decision stays on record.

313recorded changes, each with a rollback path
02 · memory

An intern on day oneA team after a year

Usually

Every morning the assistant is a new intern. It remembers no agreements, so it only gets simple tasks.

Here

The agent starts the day with the rules, project memory and the failure log. You can hand it a task for a week.

182recorded failures reduced to 11 rules
03 · organisation

A drill in the garageA crew with a site manager

Usually

A tool for single jobs: summarise an email, fix a sentence. Nothing lasting gets built.

Here

There is a plan, subcontractors and acceptance in target conditions. A human signs the handover.

748jobs for AI subagents in 44 days

Both sides often run the same engine. What decides the outcome is the organisation and the person who directs it.

Before anyone says “it can't be done”

Eight objections from the IT industry and the facts that answer them

Objection“One person cannot keep a system like that running.”
Facts24 agents and 309 sessions in 44 days. One person who decides. The rest is organisation: areas, owners, memory, a failure log.
Objection“AI must not touch infrastructure.”
FactsIt runs eight devices with me: servers, virtualisation, a private network, backups. 313 recorded changes, each with a rollback path. I type the passwords myself.
Objection“An agent will always say it works.”
FactsI know. I counted: 13 times it reported success on partial evidence. So it has two answers to choose from: checked in target conditions, or “UNVERIFIED”.
Objection“Company data must not be given to a model.”
FactsI do not give it. A local anonymiser strips it out before anything leaves the building, and the model's access to data is cut by construction, not by instruction. A human approves every draft before it is sent.
Objection“A local model is too slow and cannot be trusted.”
FactsThe same set of questions runs 4.6 times faster after choosing the right variant, with no graphics card. And trust is not required: its code runs in a sandbox with no network (6 of 6 tests), and a second model judges the result.
Objection“You won't do that in a week.”
FactsA simulator game, approx. 60k lines of C#: four days. A shared dashboard for the whole environment: two and a half hours. A customer quote with hardware selection and deployment pricing: seventeen minutes.
Objection“Away from the office you can't do anything with it anyway.”
FactsI write to the agent from a messaging app on my phone. Fourteen days of work with no break and no restart, 15 answers to 15 questions. The car reports by itself, and I can decide on a quote from the phone.
Objection“Who will maintain it afterwards?”
FactsThe person who moved a group's server room without a maintenance window. The agents have owners, scopes, a failure log and checkpoints. It is ordinary operational discipline, only faster.
One principle

Execution can be delegated. Accountability cannot.

That is how I ran infrastructure under audit and how I run agents: controlled access, a human signature, a full trail, a test before go-live.

Human

Decides

Approves, signs, sends. Higher-risk changes need consent in two steps.

AI

Prepares

Reads, selects, calculates, writes, tests, and shows where it got it from.

Behind the wall

Out of the model's reach

Personal data, passwords, money. By construction, not by request.

Where it comes from

AI stands on infrastructure. I know infrastructure from the foundations up.

A control engineer by training. I started with cables, hotel door locks and servers. I ended up running group IT through acquisitions, a company split and regulatory inspections.

16 peoplein three countries; two voluntary departures in five and a half years
3 companiesacquired in two countries and brought to one standard, with no interruption to work
0 findingsfor IT after the FDA inspection; ISO 27001 programme from 28 non-conformities to one
3 monthsto split a company into three independent companies with fully separated IT
Summary

Fourteen builds. One method of work.

Fourteen builds, each laid out step by step: what the machine does, where the human decides, what the measurements showed, what else it can do and which failures changed the way of working.