The short version
Most agencies show a demo. We can show operations. Media Express has used AI agents for daily technical work for months: moving websites between servers, setting up analytics, publishing content to many Facebook Pages, cleaning up WordPress databases and writing pages like the ones on this site. Every one of those jobs leaves a trail in a written log, and every serious mistake turned into a rule that is now enforced by a script.
The design follows what public guidance on agents recommends: keep the agent inside narrow permissions, require human approval for high-impact actions and keep logs (OWASP calls the opposite "excessive agency"), and let agents pause for human checkpoints and test them in a sandbox before real use.[2],[3]
The team: one owner and six AI models
The owner makes the decisions and clicks the buttons agents are not allowed to click. Six AI models do the work, each through its own subscription and tools, and all of them read the same shared records at the start of a session.
| Model | Typical work here | Boundary |
|---|---|---|
| Claude | Builds and repairs websites, migrations, analytics setup, scripts and reusable skills | Works through SSH, APIs and local tools; asks before outward-facing or hard-to-reverse steps |
| Codex | Code changes, WordPress and server tasks, second-opinion diagnosis | Same shared records and rules; hands off through written notes |
| Hermes | Runs scheduled jobs, a Telegram dispatcher, watchdogs and maintenance on the Szef.AI runtime | Cheap background work; alerts a person instead of improvising |
| ChatGPT | Research, drafts and marketing support, including a persistent assistant that coordinates follow-up | Cannot read our local files unless a connected task is started |
| Gemini | Research, text-to-speech and an alternative second opinion | Used for specific jobs, not for server changes |
| Meta AI (Muse) | Facebook-related research and tasks | Sees only a shared cloud folder, not our disks |
A real stretch of work: ten sites in about 30 hours
Between the afternoon of 2026-10-10 and the early hours of 2026-10-11, our agents moved more than ten websites from one hosting server to another: files, databases, mail, DNS records and SSL certificates. The owner approved the plan and did the steps that need an account holder (for example switching name servers at the registrar). The agents did the rest and proved each result before the next step.
| Site | What was checked | Logged result |
|---|---|---|
| thebest.onl | Active files compared by hash, database tables, mail restore | 15,234 files hash-verified; 121 database tables; 92 of 92 mailbox items restored |
| chicagolocal.biz | Files by hash, database tables, site verification script | 14,070 files hash-verified; 64 tables; verification passed |
| strony.us | File hashes and a URL-by-URL comparison with the old site | 129 of 129 file hashes matched; 105 of 106 URLs identical (the one difference was reviewed) |
| mediaexpress.us and its SEO tool site | Files by SHA-256 on both sites, database tables, mail records re-created | 7,268 and 3,286 files identical; 26 database tables; Google Workspace mail records re-created in the new zone |
A hash check compares a fingerprint of every file before and after the move, so a single damaged file shows up as a mismatch instead of a broken page weeks later. The same habit applies to any agent that changes systems for a client: prove the result, do not assume it.
Guardrails that exist because something went wrong
We keep a record of mistakes because they are the cheapest way to learn what an agent should never be allowed to do alone. Three examples, as logged:
Too many connections got our office blocked
On 2026-09-28 and again on 2026-10-11 the hosting firewall blocked our office address after bursts of automated SSH connections. We wrote a shared rate limiter with a circuit breaker: it spaces connections 8 seconds apart per server and, after the first timeout, pauses that server for 15 to 240 minutes. On 2026-10-12 we also found that a scheduled job was probing three servers every 15 minutes outside that limiter, and fixed it with the owner's approval.
Thirteen copies instead of a link
On 2026-09-21 an agent published 13 native copies of a Facebook post when the owner wanted traffic sent to the original Page. The agent deleted all 13 copies and added a link-share mode so the choice is explicit next time.
Backup first, always
The workspace rule is simple: before any file change in a project or site folder, a script creates a backup archive and a git checkpoint and prints its result. No printed pass means no work. This is the reason a bad edit can be undone in minutes.
These controls map to well-known mitigations for agents that can act on systems: least privilege, human approval for high-impact actions, and logging that limits damage when something still slips through.[2]
What agents may not do alone
- Grant roles or permissions (for example access to a search console property). The owner clicks; the agent supplies the direct link.
- Change name servers at the registrar, or detach a domain from the old server. The owner does it, the agent verifies afterwards.
- Delete a website, database or mailbox permanently. A backup is verified first and the owner gives an explicit order.
- Send, publish or buy on the company's behalf without a clear go from the owner (a "go" in chat counts; silence does not).
- Print or copy credentials between machines. Secrets stay where they live and agents run the call there.
Example: during a mail clean-up on 2026-10-10 an audit of 116 DNS zones found 27 missing standard mail settings. After the owner's go in chat, the agents fixed 25, left two infrastructure zones alone and re-ran the audit, which came back clean.
Content, analytics and search setup done by agents
On 2026-10-12 four AI lead sites went live in one day, each with its own project folder, analytics property with lead and phone-click events, search engine verification, a sitemap submission, an instant-indexing key and a tested mail route. The same day, twelve sub-agents rewrote twelve templated local pages for a contractor, each one with its own file of sources and a target of at most 25 percent shared text between pages.
The pages you are reading follow the same method: research with sources, a written specification, an automated check that every number has a source, and a human review before publishing.
Where you can check us independently
Our own log is our statement. These profiles are kept by third parties, so you can read what others say without taking our word for it. We do not control their content and we do not quote ratings here.
What this means for your project
Written scope first
We agree the goal, the tools, the approvals and the success measure before building anything.
Least privilege
Each agent gets its own account with only the access its job needs, on accounts you own.
People approve what matters
Money, customer messages and record changes can wait for a person, and you decide where that line is.
Logs you can read
Every action is recorded, so you can see what the agent did and switch it off.
Questions answered
Do you really use AI agents for your own work?
Which AI models do you use?
What happens when an agent makes a mistake?
Can an agent delete data or change DNS on its own?
How do you stop several agents from overwriting each other?
Why trust numbers from an internal log?
Does this mean agents can run my business unsupervised?
Is any client data used in these examples?
Sources
- AI Risk Management Framework (NIST), accessed 2026-10-12. Names the four core functions Govern, Map, Measure and Manage.https://www.nist.gov/itl/ai-risk-management-framework
- LLM06:2025 Excessive Agency (OWASP Gen AI Security Project), accessed 2026-10-12. Defines excessive agency and lists least privilege, human approval of high-impact actions, and logging as mitigations.https://genai.owasp.org/llmrisk/llm062025-excessive-agency/
- Building effective agents (Anthropic), accessed 2026-10-12. Recommends human checkpoints and extensive testing in sandboxed environments with guardrails.https://www.anthropic.com/engineering/building-effective-agents
How we checked this page
What we checked
- Dates and counts were copied from our internal change log and task board (entries from 2026-09-21 to 2026-10-12).
- The three public sources were opened and read on the access date.
What we could not verify
- The internal log is not public, so no outsider can re-check the counts.
- We have not measured business results (revenue, savings or rankings) for any of the work shown.
Want agents run this carefully for your company?
Describe one process. We reply with a written scope that names the tools, the approvals and the success measure.