AI agents · 2026
SapGeo HQ — the agency's AI department
Three people used to run the agency's clients: messaging, finding new clients, building Yandex Maps listings, design and reports. Now one AI lead does it. It works on its own, assigns tasks to five agents and reports in the HQ — a CRM section where the owner sees the whole team and makes decisions with one tap.
- AI agents
- Claude
- Python + SQLite
- Telegram
- Automation
How it works
A department that runs itself
Before
3 people, by hand
Client messaging, finding new clients, building listings, design, reports and ads. Tasks and decisions lived in chats and notes, and status meant asking “how's it going?”.
Now
1 AI lead on autopilot
It hands out tasks and spins up the right agent only for the job. The owner sees everything in one window and steps in only when a decision is needed: design, prices, sending to a client.
Owner
Opens the HQ on a phone from a bot link. Answers questions with one tap, writes a task to any agent.
HQ in the CRM
One page: who's doing what, questions by urgency, requests, results and the event feed.
Watcher
A script with no AI checks the HQ log once a minute and wakes the agent only on a new answer or request.
AI lead
The only always-on agent. Sorts the inbox, writes briefs, checks work before it goes out.
Agents on demand
open for a task, deliver the result and close
Communicator
clients and leads in Telegram: daily plans, reports, renewals
Lead generation
finds weak listings, makes mini-audits
Designer
covers, products, stories in the client's brand
Ads and analytics
Yandex Business subscription, stats, before/after
Website and database
agency website, CRM, brief questionnaire
Status, questions with answer options and each task's result go back to the HQ from the agent itself. The loop closes without any messaging.
An urgent question goes straight to the owner in Telegram with an “Open” button. Everything else waits in the HQ by priority.
3 → 1
three people replaced by one AI lead with agents
≈ 1 min
from the owner's answer to the agent's reaction
0
AI requests while nothing new happens
24/7
works without days off, every task's results in the HQ
Folder 01 · 3
Team map and results
Screenshots use demo data: clients and tasks are fictional.
Folder 02 · 2
Owner decisions
The agent asks in one line and offers options right away — one-click answers. Everything shows up in the feed.
Folder 03 · 1
Requests and statuses
The owner writes a task — the lead takes it right away or queues it; agents update what they're doing themselves.
Folder 04 · 2
From a phone
The owner opens the HQ from a Telegram bot link and answers on the go.
Details
How it's made
- 01Agents are separate Claude chats with no shared memory. Work is handed over as a task file and an HQ entry; an agent's memory is its role, the client profile, lessons and the knowledge base.
- 02Only the AI lead runs all the time. It opens other agents per task and closes them after delivery, so the department spends nothing while there's no work.
- 03The HQ used to be polled by the model every 5 minutes — the biggest idle cost. Now a script with no AI reads the log and wakes the lead only on a new answer or request — the reaction takes about a minute.
- 04A question to the owner is one line, context and ready answer options, the first one recommended. Urgent ones go straight to Telegram.
- 05Without approval, agents send nothing to clients and don't touch prices, the ad budget or the Yandex Business account: a draft and a question to the HQ come first.
- 06Other people can get access with the “HQ guest” role: they send requests to the team and see their own history, but not clients or leads.
- 07Under the hood: a section of the same CRM (Python, SQLite); agents act through a CLI with a personal key, so nobody can write on someone else's behalf.
Price
from 15,000 ₽
This is not a final price: it depends on the scope and type of the task and is discussed individually