AI agents that run part of the company
At Reefa and KtoSprząta.pl I built 11 AI routines and agents, and 10 of them run every day. They check the panel with 44 database queries, fix and deploy code, develop the websites, moderate content, look for tenders and send me a brief every morning. Since 16 October five of them decide on their own and append every decision to a log that can only be added to. What they mostly leave to me is logging in, paying, signing and disputed cases.
This makes sense in a company where someone does the same checks, reports and small fixes by hand every day. AI agent projects start from PLN 15,000 excl. VAT. I am based in Kraków and work remotely and on site across Poland.
11
AI routines and agents I built; 10 run every day (as of 16 Oct 2026)
- Company
- Reefa and KtoSprząta.pl
- Period
- June–October 2026
- My role
- Owner. I write the rules, build the routines and read their logs.
- A similar project
- from PLN 15,000excl. VAT
What it looks like on screen
| Start | Routine | What it does |
|---|---|---|
| 5:40 | Panel doctor44 database checks and panel fixes | 44 database checks and panel fixes |
| 6:45 | Morning briefup to 20 lines in Telegram | up to 20 lines in Telegram |
| 7:07 | KtoSprząta.pl blogarticle of the day with a cover | article of the day with a cover |
| 7:15 | Lead growthchanges and form tests on reefa.pl | changes and form tests on reefa.pl |
| 8:03 | TendersBZP and TED, scoring and offer pack | BZP and TED, scoring and offer pack |
| 8:20 | KtoSprząta.pl moderationalso at 14:20 and 20:20 | also at 14:20 and 20:20 |
| 11:00 | KtoSprząta.pl developmentone backlog item | one backlog item |
| 13:00 | SEO contenttitles and copy on reefa.pl | titles and copy on reefa.pl |
each of these runs on its own
- Date
- date and time
- Question
- the question the agent settles
- Options
- options A/B/C
- Choice
- the chosen option
- Why
- reasoning
- How to roll back
- rollback command
Append-only
Why a company needs agents, not another chat window
An AI chat helps when someone sits down and asks a question. But most of this kind of work in a company is repetitive checks and small fixes that nobody has time for every day: is there an error in the data, do the website forms work, is there a new tender, has someone posted content that needs moderation.
I wanted these things to run on their own, but under control. An agent that can deploy code and touch data needs hard limits, has to record its decisions and must be able to roll itself back. Without that I wouldn’t trust it with even a website.
What has to be true before you hand work to an agent
- You know what it decided and why, without reading code
- Any change can be undone in seconds
- It never touches money without a document or a bank record
- It can’t change its own rules
- You only get what the agent physically can’t do
What I built
Safeguards and the decision log
Since 16 October five routines (lead growth, SEO content, the panel doctor, and KtoSprząta.pl development and moderation) decide for themselves and append every decision to a log: date, question, options, choice, reasoning and how to roll back. The rules live in locked files with a SHA-256 checksum, so an agent can’t change them without my approval. Reports have to be readable in 20 seconds.
Panel doctor
Every day at 5:40 it runs 44 database checks (money, working time, operations) across 20 areas, fixes the panel frontend itself and deploys it. 69 money-related files fall under stricter rules: financial data may change only on an exact match with a bank record or a document, and a new formula ships only after a reviewer recalculates at least two examples by hand.
runs on its ownLead growth on reefa.pl
Daily at 7:15 since 4 October. A full autopilot for the website including deployment, only through its own ship script and only in the website repository. It sends a test enquiry through every form and takes screenshots. I measured rollback at 3 seconds and switching to the new version at 2 seconds. By 16 October it had logged 33 decisions.
runs on its own
SEO content on reefa.pl
runs on its ownDaily at 13:00 since 14 October. It changes any one URL template at most once every 28 days and judges the effect after 35 days (at least 0.4 percentage points of CTR at 500 impressions). A content guard blocks competitor names, unverified ratings, site counts and copied passages of 8 words or more.
KtoSprząta.pl moderation
runs on its own3 times a day. It edits only by striking out whole words, and any fix or rejection needs a second agent to agree. Limits: 30 decisions and 10 rejections per run. It never edits reviews, authors get a statement of reasons under Article 17 of the DSA, and legally disputed cases come to me.
KtoSprząta.pl development
runs on its ownEvery day at 11:00 it ships one backlog item, with automatic rollback. Database changes may only add, and protected zones (keys, login, the database folder) never ship.
Tenders
runs on its ownEvery day at 8:03 it scans the Polish public procurement bulletin (BZP) and TED, reads the tender documents, looks up past results and scores each one: go, watch or no-go. A “Prepare offer” button assembles a 5-document package: offer form, statutory declaration, service list, costing and filing instructions.
Morning brief
runs on its ownEvery morning (6:45 in summer) I get at most 20 lines in Telegram: money, sales, operations, people, up to 3 decisions for the day and the main risk. This brief deliberately runs on a fixed template, with no language model.
Company search
It searched the company register (KRS), the schools register (RSPO), the VAT white list and company websites, scored companies 0–100 and stored the legal basis with every record. I paused it on 25 August, with about 7,100 companies already in the queue.
First contact
runs on its ownThe first email only asks for consent to send an offer, and the agent sends the offer only after a “yes”. A server-side gate (suppression list, no repeats, daily cap) works independently of the agent. When a reply contains a phone number, the agent stops writing and I call.
reefa.pl blog
runs on its ownOne article a day prepared with Claude, with backup runs in case of an error. The company facts are written into the instructions, so the model doesn’t make them up.
KtoSprząta.pl blog
runs on its ownAn article every day at 7:07, with a cover image picked by a vision model. The routine refills its own topic queue.
How it fits together
1. Start
Schedule
at a fixed time
runs on its own
2. Agent
Playbook with red lines
locked, SHA-256 checksum
Decision
appended to the log
runs on its own
3. Checks
Review from three angles
runs on its ownTypes, build and tests
runs on its ownMoney only on a match
with the bank or a document
runs on its own
4. Release
Test build on a separate address
runs on its ownHas anyone else changed the live version?
runs on its ownSwitch to the new version and test live
runs on its ownRoll back in 3 s
if the test fails
runs on its own
5. Owner
20-second report
in Telegram
a person decidesLog in, pay, sign
and disputed cases
a person decides
A routine starts on schedule, decides by its playbook and logs the decision, passes review and tests, deploys with a live-version check and automatic rollback, and leaves me a short report.
Key:runs on its owna person decidesdata, people or a place of work
Numbers from the system
| What I measure | Value | As of |
|---|---|---|
| Routines and agentsas of 16 Oct 2026 | 11, one of them paused | |
| Panel doctor checksas of 13 Oct 2026 | 44 queries across 20 areas | |
| Money files under stricter rulesas of 13 Oct 2026 | 69 | |
| Website rollback timeas of 4 Oct 2026 | 3 s (switching to the new version 2 s) | |
| Decisions logged by the growth routineas of 16 Oct 2026 | 33 | |
| Queries tracked by the content routineas of 14 Oct 2026 | 32 in Kraków, 16 in Katowice | |
| Questions in the AI visibility checkfirst measurement scheduled for 22 October 2026as of 14 Oct 2026 | 14, in ChatGPT, Gemini and AI Overviews | |
| Reviewer findings before the content routine went liveas of 14 Oct 2026 | 30 from 3 reviews, all applied | |
| TED tenders added in one runas of 13 Jun 2026 | 21 from Małopolska and Silesia | |
| Morning brief lengthas of 12 Jun 2026 | up to 20 lines, no language model |
The brief, the tender agent and the sales agents were set up in June; the panel doctor, the content and growth routines and KtoSprząta development and moderation in October. I haven’t measured their effect on enquiries yet, so I don’t quote one.
Besides these 11 there are smaller scheduled jobs: a monthly cash review, client reports on the 5th, a daily month-close check and a quarterly price and law review for KtoSprząta.pl.
What it runs on
- Anthropic Claude
- agents, reviewers and decisions
- Schedulers and pg_cron
- starts at fixed times
- Vercel
- test builds, switching versions and rollback
- Supabase: Postgres, Edge Functions
- data, SQL checks, server-side gates
- Telegram Bot API
- my reports
- BZP, TED, KRS, VAT white list
- sources for the tender and sales agents
- Google Search Console
- measurement for the content routine
What you can take for your company
Start with checks, not deployments
Your first agent should check, not change: daily data checks and a morning report. It earns the right to make changes only once you know how it gets things wrong.
A decision log instead of trust
Every decision is recorded with the reasoning and a way to undo it. That way you know what happened in 20 seconds and don’t have to watch the agent all day.
Money only on hard evidence
An agent may propose, but it changes financial data only on an exact match with the bank or a document, and after a second opinion. I apply this rule wherever AI touches money.
Not everything needs AI
My morning brief runs on a fixed template with no language model, because it has to be the same every day and always match the data. I pick the tool for the job, not the other way round.
Services where I do this
AI assistants and agents that read documents, answer customers and draft offers — connected to your data, with a human in control.
from PLN 15,000excl. VAT- from PLN 15,000excl. VAT
Other case studies
- 13An operating system for a service companyOne panel for the whole company: schedules and shifts, finances across several entities, petty cash, bank matching, e-signed contracts and client reports.
- 1,100 hA bot for the field teamA shift starts and ends with one button in Telegram, with location and photos. The rebuild removed 55 issues and is covered by 19 scenario tests.
- +30%Organic growth of reefa.plA plan of 33 SEO and GEO changes, 29 free tools, an open norms dataset (460 categories from 2,338 rows of data) and a content routine. Clicks from Google rose from 8.6 to 11.2 a day in July.
- 3 daysKtoSprząta.pl — an industry platformA directory of companies from public registries, a search that fixes typos, a free Pro panel for companies and AI moderation three times a day.
Let’s talk about your company
30 minutes, no sales deck. You tell me where time and money leak, and I tell you what I’d do first and roughly what it costs.