СВОЙ Тех builds an internal AI agent platform: 200 tools and 5 chats

The team at СВОЙ Тех, a fintech company, has published a breakdown of its internal AI agent platform on Habr. The title names two numbers: about 200 tools and 5 chats. The authors are building their own AI assistant inside the corporate infrastructure, not a layer on top of a third-party API. The announcement lists three topics: the architecture "under the hood," the set of connected tools, and context management. The analysis itself is in the full article, so below we cover only what the announcement lets us say with confidence, plus the engineering logic around it.

Prototype vs. platform

The authors' starting point is simple: nobody is surprised by AI integration into workflows anymore, but the gap between a quick prototype and a system inside a company is huge. A prototype on an external API comes together in a couple of days. It answers questions, writes drafts and looks good in a demo. But once the agent needs internal data, access rights and real business processes, a different kind of work begins: integrations, data isolation, control over actions, reproducibility of answers.

The numbers in the title point to where the main difficulty lies. Two hundred tools means two hundred function descriptions, parameter schemas and error scenarios. If you hand them all to the model at once, the descriptions eat a noticeable share of the context window, and choosing the right call becomes less reliable. The announcement does not reveal how the five chats are divided. But the logic is easy to read: a chat for a class of tasks narrows the set of available tools and data, and with it the space of possible errors.

What the announcement tells us

The teaser contains few facts, but they are concrete:

• Developer: СВОЙ Тех, fintech division.
• Format: an in-house AI assistant and an internal platform of specialized agents.
• Scale of integration: around 200 tools.
• Interface: 5 chats, each for its own business tasks.
• Environment: corporate infrastructure, working with internal data.
• A separate section of the article: context management.

The last item is arguably the most interesting. In agent systems, context is a limited resource. Dialogue history, tool call results, document excerpts and system instructions all share it. Who decides what to keep and what to evict or compress, and how, often shapes the quality of the system more than the choice of model does.

What to check if you are building something similar

Until more details appear, the announcement works as a checklist of questions for your own project. Ask them before opening the full article.

1. How many tools does the model see in a single request? If it sees all of them at once, it makes sense to group them by task or load them on demand.

2. Is there a split by roles and scenarios? Several specialized chats are easier to test and to restrict by permissions than one universal chat.

3. How is context handled? You need explicit rules: what is stored, what is compressed, what is passed between agents.

4. Where does the data live? For a fintech, whether it stays inside the perimeter determines the entire architecture.

5. How is the result measured? Without a set of test tasks and quality metrics, a growing number of tools easily turns into growing chaos.

What matters most in the СВОЙ Тех publication is not the size of the platform but that the company shows its engineering kitchen: how a set of integrations becomes a working system. How convincing the details are will become clear from the full text, but the direction is clear. Corporate AI agents in the financial sector are moving away from a universal chat toward a set of narrow, controlled perimeters with their own context management.