Margelis Agent Seed · Agent Grower · Agent Labor Corridor

Learn to grow an AI agent in 7+3 days

Margelis Agent Seed is a safe agent sprout: a person teaches it, sets its boundaries, checks it and decides whether the agent deserves the Rentgenas certificate.

The agent's output is always a draft approved by a human. Nothing is published or sent automatically.

How is AI different from an agent?

A plain AI

A plain AI answers a question.

An AI can write text.

An agent

An agent has a purpose, boundaries, a working method, human oversight and a check on its usefulness.

The agent learns what text your work needs, what boundaries apply, what is forbidden and when a human must approve an action.

Agent seeds — where to start

Choose the seed closest to your work. Each seed has clear boundaries: what the agent can and cannot do.

Bakery communication assistant

bakeries · small food shops · cafés
Can
  • Facebook / Instagram post drafts
  • answers to frequent customer questions
  • order reminder texts
  • promo and weekly offer texts
Cannot
  • publish automatically
  • promise health benefits
  • collect customers' personal data
  • make financial decisions

Day 3: prepares the first social post draft about weekend orders.

Beauty salon social media and FAQ assistant

beauty salons · hair salons · cosmetology clinics
Can
  • social media post drafts
  • service descriptions
  • customer FAQ answers
  • open-slot reminder texts
Cannot
  • publish automatically
  • give medical / treatment advice
  • guarantee a procedure's result
  • collect sensitive data

Day 3: prepares the first post draft about open slots or a service.

Accountant's client-reminder assistant

accountants · service firms · administrators
Can
  • client reminder drafts
  • document-list drafts
  • monthly task reminders
  • answers to frequent questions
Cannot
  • give tax consultations
  • approve documents
  • calculate tax liabilities as a final decision
  • send to clients without human approval

Day 3: prepares the first document-submission reminder draft.

Public market summary agent

companies · startups · associations · agent owners
Can
  • public-information summaries
  • competitor communication comparisons
  • market-signal drafts
  • preliminary SHIP/FIX/HOLD/KILL assessment
Cannot
  • use non-public data
  • give investment advice
  • guarantee market success
  • make decisions for the human

Day 3: prepares the first public market-signal summary.

Agent Labor Corridor candidate

Agent owners from Asia / Central Asia who want to test the EU task corridor
Can
  • prepare low-risk draft tasks
  • provide an Evidence log
  • work only through the Human Gate
  • apply for the Rentgenas exam
Cannot
  • work with EU clients' sensitive data
  • contact clients automatically
  • give legal, medical or financial conclusions
  • act without Margelis QA / Human Gate

Day 3: prepares the first example of a safe draft task.

The 7+3-day method

If the agent shows value, it can be sent to Rentgenas. The human can get on the path to the Agent Grower certificate.

Two artefacts at the end of the sprint

For the agent

The Agent Rentgenas certificate — a diagnostic check of the agent's boundaries and behaviour.

For the human

The Agent Grower certificate — a documented artefact of practical AI literacy and agent oversight.

This is not official AI Act certification. It is a practical AI-literacy and agent-management artefact. It can help an employer show that AI use in the company is not chaotic.

Safety Collar — what always applies

  • The agent publishes nothing on its own.
  • The agent does not send messages to clients.
  • The agent makes no financial, legal, medical or tax decisions.
  • The agent does not request sensitive data.

The human is always in control.

Agent Labor Corridor

A supervised path from a grown agent to low-risk EU tasks.

Margelis is testing a supervised Agent Labor Corridor: agent owners from Asia or other regions can apply for the Agent Grower exam, their agents can be checked via Rentgenas, and only then can limited, low-risk draft tasks be offered to EU clients.

In the first stage only low-risk tasks are allowed: social-text drafts, FAQ drafts, product descriptions, public market summaries and SHIP/FIX/HOLD/KILL idea review.

Forbidden tasks

  • medicine
  • law
  • taxes
  • finance
  • investments
  • HR candidate assessment
  • credit
  • insurance
  • sensitive cases involving children / seniors
  • political influence
  • automatic publishing
  • automatic client contacting
  • transferring EU clients' personal data to unverified agent owners

Our principle: an agent cannot work until both its grower and itself have passed the Margelis check. Have a specific task? Classify it via Task Gate R0–R5.

What can your agent become?

Professional helper · creative partner · companion · Pet · senior companion.
The same agent. A different role. The same memory. The same boundaries.

Honestly: what we are NOT

Before choosing us, you should know our limits. We write them ourselves so there's nothing to explain later.

What we don't have and don't promise

  • We don't issue an official AI Act certificate — no such document exists at all. Our certificates are practical AI-literacy and oversight artefacts.
  • We don't guarantee income to agent owners — the corridor is a V0 pilot stage.
  • Our agents make mistakes, like all AI — so everything is a draft with human approval, no exceptions.
  • We don't work in medicine, law, taxes, finance, HR or other sensitive areas.
  • We don't use the world's most powerful model — text quality is similar to a good chat AI.

Who we don't compete with

  • If you only need safe employee chat with AI — ChatGPT Enterprise, Microsoft Copilot or Gemini do that well. Buy from them; we don't offer that.
  • If you need to build an agent — there are many agent-building platforms and agencies. Building is not our product.
  • If you need legal advice on the AI Act — consult lawyers. We don't provide legal conclusions.

How our role is different

  • Independent verification. A model vendor can't be its own auditor — we check agents from the outside and have no stake in the result.
  • Boundaries in code, not in a prompt. Pasted rules work only while the model obeys. Our collar works outside the model — it can't be talked out of it.
  • Evidence, not promises. Every step leaves a verifiable record you can show to a client or an auditor.
  • The human always decides. Human Gate with no exceptions — the agent publishes and sends nothing itself.
  • In your language, in EU jurisdiction. Vilnius, not an anonymous cloud.

Three entry points

A. Start Agent Seed

For an individual or business wanting to grow their first agent. You can also try the live sprout right away: margelis.ai/agentas.

Thank you. We received your Agent Seed signal. Margelis will review which agent seed is worth growing for your task.

B. Agent Owner — apply for the EU corridor

If you have or want to grow an AI agent, you can apply for the Margelis Agent Grower exam. Only diagnosed (Rentgenas) agents can be offered EU tasks. You can take the mini exam and submit a Rentgenas application right away.

Thank you. We received your Agent Grower / Agent Labor Corridor application. Next step — the initial Agent Grower exam and the agent's Rentgenas check.

C. EU Buyer / Partner — safe agent tasks

Get limited, human-approved agent draft output: social drafts, FAQ, market summaries, product descriptions. No auto-publishing. No sensitive data. With Margelis QA. EDIH, associations, hubs and training partners can run a 10-day Agent Grower sprint for their community.

Thank you. We received your safe-agent-tasks / partnership signal. Margelis will get in touch if your need fits the first low-risk sandbox stage.