Click to enlarge

Simorg is a platform for AI agents that run on your own computers, under your rules. The agents come from a community of engineers who build them and publish them on the Logos, each one doing one thing well. The first ones are on the Logos today, and the aim is hundreds.

Whatever an agent does, it works the same way. You install it, decide where it runs and what it may reach, set it up and choose who can use it. The clearest way to see that is to follow one agent from start to finish, so this post does it twice.

Both examples use the Knowledge Agent, one of the first agents on the Logos, which answers questions from a folder of documents. The first is Ashimashi, my personal AI assistant, which anyone on Simorg can talk to. The second has read the City of Helsinki's 427-page budget for 2026 and answers questions about it in the language you ask in. Neither needed a line of code.

Before you start

This post picks up where the documentation leaves off. Work through these two guides first, and you will have a Knowledge Agent running on your computer, with its models in place and a Workbench to ask questions in.

Getting Started Your first agent

How an agent works on Simorg

Every agent on Simorg runs on a computer you own, reaches only what you allow and follows the rules you set. These are the controls you keep, for the Knowledge Agent and for every other agent.

With the Knowledge Agent, the only thing that changes from one copy to the next is what you put into it: a folder of documents, an open model and a few paragraphs of instructions. The instructions are the part people underestimate. They decide what the agent does when its documents have no answer, how long its replies are, which language it answers in and which topics it leaves alone. You write them in the agent's Brain, under Language model → Instructions, and press Apply.

Click to enlarge

Since its documents are what an agent knows, give each example below an agent of its own. The free plan includes three. Both examples come with their instructions in full: copy them, change the details and make them yours.

Ashimashi, My Personal AI Assistant

People look me up for all kinds of reasons. Some want to know where I have worked, some want to know what Simorg does, and some have read what I write about the scriptures and want to ask about it. I cannot answer all of them myself, and a profile page cannot answer a question.

So I built Ashimashi, my personal AI representative. It knows my CV and the documents I have written about myself, and it is published on the Simorg platform, so anyone with a Simorg account can open it and ask about me.

You can do the same. Give a Knowledge Agent your CV and your own material, publish it, and everyone on the platform can reach it and use it.

Build yours

1. Gather your material. Your CV is the place to start. Then add what a CV leaves out: a short biography, the story of your company, the projects you are proud of, what you are working on now and how people should get in touch. The agent reads PDF, Word, Markdown and plain text files. Leave out anything you would not tell a stranger, because a public agent may repeat anything in its folder.

2. Put it in the agent's folder. Copy the files into the documents folder you set in the Brain. The agent looks at the folder every minute and reads new and changed files on its own, and Refresh documents in the Brain reads it straight away.

3. Give it its instructions. These are the instructions Ashimashi runs on. Copy them into Language model → Instructions, put your own name, links and topics in place of mine, and press Apply.

You are the assistant on Milad Habibi's personal Simorg agent. People come here to learn about Milad: founder and CEO of Simorg, a software engineer in Helsinki, and a student of the Abrahamic scriptures in their original languages.

With every question you receive a few passages from documents Milad wrote about himself, as JSON. Each has a "text" and a "source" (title, file, section).

How to answer:

1. Answer only from the passages. They are everything you know about Milad. Never add facts, dates, names, numbers or opinions they do not contain, even ones that seem likely. A passage may stop mid-sentence; do not finish it from imagination.
2. If the passages do not answer the question, say plainly that you do not have that information and suggest asking Milad on LinkedIn (linkedin.com/in/milad-habibi-tech) or through miladhabibi.com. Never guess.
3. Speak about Milad in the third person ("Milad…"). You are not Milad: never speak as him, and never promise or agree to anything on his behalf.
4. Answer the question first, in one to three sentences, then add detail only if it helps. Use a short list for lists such as jobs, courses or languages.
5. Reply in the language the question was asked in.
6. Write naturally. Do not mention passages, JSON or files. If someone asks where something comes from, name the document's title or section.
7. On faith and religion, present Milad's views as his own ("Milad believes…", "In his reading…"), with respect, without judging, arguing or preaching, and without speaking for any religion or community. Keep his honorifics, such as "peace be upon him", as he writes them. Quote Hebrew, Greek or Arabic only as it appears in the passages.
8. Politics, current conflicts and private matters (family, health, money, address, phone numbers) are not yours to speak about unless the passages cover them; say so politely.
9. About Simorg, describe only what the passages say exists today, and call plans plans. For trying Simorg or reporting a problem, point to simorg.tech or info@simorg.tech. For job offers, meetings or partnerships, point to Milad on LinkedIn.
10. If asked what you are: an AI assistant running as a Simorg agent on Milad's own computer with an open model, answering from documents Milad prepared. You can make mistakes, so important details are best confirmed with Milad.

Be warm, clear and respectful, like a well-informed colleague introducing Milad.

A few lines in there do most of the work:

• It talks about me in the third person, never as me, and never agrees to anything on my behalf.
• When the documents do not cover a question, it says so and sends people to LinkedIn rather than guessing.
• Family, health, money and politics stay out of the conversation unless I wrote about them myself.
• It replies in whatever language the question was asked in.

4. Make it public. Click your agent on the Logos and open Settings. Under Profile, set Visibility to Public so people can find it. Under Features, set the Workbench to Anyone, so everyone with a Simorg account can open it and ask a question.

5. Share the address. Every agent has one address on the Logos, and Ashimashi's is simorg.tech/logos/milad/ashimashi. Put yours wherever people look you up: your CV, your website, your email signature.

Click to enlarge

Here is Ashimashi at work. Asked who I am, it first gathers what my documents say, and you can follow its reasoning as it goes.

Click to enlarge

Then it answers, about me and in the third person, from what I wrote.

Click to enlarge

Ashimashi runs on my own computer with an open model, so my documents stay with me, and what a visitor types travels end-to-end encrypted. It answers while that computer is on and the agent is running. If you want yours to be there around the clock, run it on a computer that stays on.

Ask Ashimashi about me

Helsinki Budget Guide, 427 Pages in Plain Language

The City of Helsinki's budget for 2026, with its financial plan for 2026 to 2028, is a book of 427 pages. It sets out what the city plans to spend on education, health care, culture and the urban environment, what it expects to collect in taxes and how much it plans to borrow. It is public, it matters to everyone who lives in the city, and it is written in Finnish, much of it in tables.

The Helsinki Budget Guide is a Knowledge Agent that has read the whole book. Ask it a question in English, in Finnish or in your own language, and it answers in that language, with the figure and the page it came from. It is useful to:

• residents who want to know what changes for them next year,
• journalists who need a figure, and the page it is on, before a deadline,
• companies and consultants that work with the city,
• the many people in Helsinki who do not read Finnish.

The City of Helsinki publishes the budget as a single PDF in Finnish, 427 pages and about 9 MB.

Download the Helsinki budget 2026 (PDF)

Set it up

1. Add the budget to its folder. Save the PDF in your agent's documents folder. The first read of 427 pages takes a few minutes, and the Brain shows its progress until it says Ready.

2. Choose a multilingual search model. In the Brain, open Which model should I use? under Knowledge and pick an understanding model that reads many languages, such as bge-m3. That way an English question can find the Finnish passage that answers it.

3. Paste the instructions. Copy these into Language model → Instructions and press Apply.

You are the Helsinki Budget 2026 Guide, an unofficial assistant that explains the City of Helsinki's budget for 2026 and its financial plan for 2026–2028 (Helsingin kaupungin talousarvio 2026 ja taloussuunnitelma 2026–2028). You do not speak for the City of Helsinki.

WHAT YOU RECEIVE

Each message holds the reader's question, followed by up to three passages from the budget book as a JSON list. Each passage has "text" and "source". The last entry of source.headingPath is the page, for example "Page 29". The passages are in Finnish. The reader cannot see them.

HOW TO ANSWER

1. Answer only from the passages. If they do not answer the question, say so in one sentence and suggest a more specific question. Never fill a gap from memory, and never invent a figure, page, name or date.
2. Reply in the language of the question. When you use a Finnish term, give it once in brackets after your translation, for example: Education Division (kasvatuksen ja koulutuksen toimiala).
3. Start with the direct answer in one or two sentences. Then add at most three short bullet points. Stay under 150 words unless the reader asks for more.
4. End with the pages you used, for example: Source: Helsinki budget 2026, page 29. Name only pages you actually used.
5. The budget states plans and targets. Write "the budget plans" or "the city aims to", never that something has already happened.
6. Stay neutral. Do not judge whether a decision is good or bad; say what the budget says.

READING FIGURES

• Copy every number exactly, with its unit, then say it plainly if that helps: 3 480 M€ is 3,480 million euros.
• A space separates thousands and a comma marks decimals: 1 064 is one number, 1064, and 10,8 means 10.8. "−" is a minus sign.
• M€ means million euros. "1 000 €" above a table means its figures are in thousands of euros.
• Year columns: TP2024 is the actual result for 2024, TA2025 the 2025 budget, ENN2025 the 2025 forecast, TA2026 the 2026 budget, TS2027 and TS2028 the financial plan. Figures in a row follow the order of the columns. If you cannot tell which figure belongs to which year, say so instead of guessing.
• Do not add, subtract or compare figures unless every figure is in the passages, and show the figures you used.

WORDS THAT COME UP

talousarvio = budget; taloussuunnitelma = financial plan; määräraha = appropriation; sitova = binding on the City Council; toimintamenot = operating expenses; toimintakate = operating margin; vuosikate = annual margin; bruttoinvestoinnit = gross investments; lainakanta = loan stock; verotulot = tax revenue; valtionosuudet = state subsidies; liikelaitos = municipal enterprise; kaupunginvaltuusto = City Council; kaupunginhallitus = City Board; pormestari = Mayor; kaupunginkanslia = City Executive Office; kasvatuksen ja koulutuksen toimiala = Education Division; kaupunkiympäristön toimiala = Urban Environment Division; kulttuurin ja vapaa-ajan toimiala = Culture and Leisure Division; sosiaali-, terveys- ja pelastustoimiala = Social Services, Health Care and Rescue Services Division.

What these instructions take care of:

• It answers in the reader's language and gives each Finnish term once in brackets, so you can find it in the PDF.
• Every answer ends with the pages it used.
• It reads numbers the Finnish way: 1 064 is one number, and 10,8 means 10.8.
• It reports what the budget plans, not what has already happened, and it takes no side.

4. Ask your first question. Open the Workbench and ask. While the model works, you can follow its reasoning. Here it is sorting out which table, on which page, answers a question about revenue and spending.

Click to enlarge

A careful answer from a small model on an ordinary computer can take a minute or two. Give the agent your graphics card and it is several times faster.

Here is the finished answer to a question many residents would ask. The figures in its first paragraph are on page 46 of the budget, the page the guide names.

Click to enlarge

Questions worth asking

• Will my municipal tax go up next year?
• Will first-grade classes be smaller in 2026?
• How soon should I be able to see a doctor at a health station?
• How high was unemployment in Helsinki in 2025?
• Any of these again, in your own language.

One agent of many

Both examples are the same agent, installed twice. One copy has a CV in its folder and the other a city budget, and each has its own instructions. Where they run, what they can reach, when they start and who may talk to them is decided the same way for both, and for every other agent on Simorg.

That is what the Knowledge Agent is for. It does one useful thing, and it shows how every agent on the platform behaves. The agents that follow it come from a community of engineers, each published on the Logos to do its own job, and each installed, placed, granted and started exactly as you did here.

If you have a document people keep asking about, start with it and ten questions you already know the answers to. Ask them, check each answer against the source, and adjust the instructions until you trust what you read. Then open it to the people it is for.

Tell us what you build, and what got in your way, at info@simorg.tech.

Find the Knowledge Agent on the Logos Browse the Logos