Our AI agents, run from a phone.
We run our studio with AI agents that work on our servers around the clock. We give them tasks in Telegram, they report back there, and a small office of agent roles handles our blog, reports and order delivery on a schedule. This is the system behind it.
In-house system · not client work
- Bots in production
- 5 Telegram bridges
- Self-healing
- 2-minute watchdog
- Open source
- tg-relay-watchdog (MIT)
- Agent roles
- 8, with a QA gate
The problem
AI agents are only useful if you can reach them
An AI agent on a server can write code, publish content and check numbers, but it lives in a terminal. We wanted to hand out work and read results from a phone, at any hour, without opening a laptop. The first version worked, then failed in quiet ways: messages arrived only sometimes, or stopped arriving while everything looked healthy. A bridge that fails silently is worse than none, because you think the work is moving when it isn't.
What we built
1. A Telegram relay that heals itself
Each agent has its own Telegram bot. A small relay receives messages for that bot and types them into the agent's terminal session; the agent answers through a send script. Five of these bridges run in production, one per project.
What makes it reliable
- Exactly one poller. Two pollers on one bot token make Telegram split messages between them, which looks like messages "arriving sometimes". The watchdog counts pollers and kills strays, keeping a single one.
- Watchdog every 2 minutes. A process manager only notices a crash. On our main bridge the watchdog also restarts the relay when pending messages stop draining, recreates the agent's terminal session after a reboot and starts a stopped service.
- Allow-list. Only approved chat IDs can send input.
- Screenshots and voice. Photos sent in Telegram are downloaded and handed to the agent as a file path, and voice messages are transcribed to text with a speech-to-text model.
- Replies that confirm. The send script checks Telegram's answer and retries instead of failing silently.
We then published a clean, generic rewrite as open source: github.com/vladosskins178/tg-relay-watchdog. It folds those lessons into the relay itself: a file lock so a second copy exits at once, the update offset saved to disk so a restart neither loses nor replays messages, and configuration through environment variables. It is Python standard library plus small shell scripts, MIT-licensed, with no webhook, open port or domain needed.
2. An office of AI agent roles
On top of the bridges runs what we call the office: a set of agent roles with written briefs, one shared task board and fixed wake-up times. Each role has a goal, inputs, outputs and a checklist.
- Task boardEvery task moves through inbox, doing, review and done. Agents update it when they start or finish anything.
- Eight rolesScout, reels, orders, support, analyst, free traffic, paid ads and QA, each with its own brief.
- ScheduleA daily morning run for numbers, orders and the support inbox; a progress check every 30 minutes that pushes stalled work; a weekly review; blog runs three times a week.
- Quality gateNothing reaches the owner or the public without an independent check: the QA role, a review score of at least 8.5 out of 10, or the role's checklist for routine items.
- Human decisionsSpending money, opening accounts and publishing under the owner's name always need the owner's approval. The agents decide everything else.
What the office runs today
- The SEO blog on this site. Each run starts with a demand check (Search Console queries and what people are asking this week), then writing, a fact-check against primary sources and a reviewer score of at least 85 out of 100. Publishing is gated by automated build checks for broken links, structured data, the sitemap and prices that must match one price file. Then the new pages are submitted to search engines.
- Daily analytics. The analyst role collects store orders, search data, email delivery logs and website leads each morning, and messages the owner only when something changed or needs action.
- Gift order delivery. The orders role watches the gift game pipeline: photo intake, 3D hero approval, publishing and the email with printable QR cards.
How it is built
This page describes our own internal setup at a high level. Server details, tokens and account names are left out on purpose.
We can build this for you
The same patterns run customer-facing work: a Telegram or WhatsApp bot that never silently drops a message, automations that check their own health, and AI agents with clear roles, a human approval step and a quality gate. Chatbots start from $2,000, workflow automation from $2,500 and custom AI agents from $6,000, with a fixed quote before we start.