The cleanest Telegram and AI-agent workflow uses Telegram as a fast human surface and a shared task board as the source of coordination state. Messages can create or discuss work; agents can execute and update it through MCP; both sides read the same task rather than maintaining separate lists.
The five-step loop
- Capture: a teammate creates a task from Telegram with the desired outcome.
- Triage: the team adds board, priority, owner, and acceptance criteria.
- Execute: an MCP-connected agent claims the task and works from linked evidence.
- Review: the agent posts results and moves the task to review; a human decides.
- Close: validation is recorded and the task moves to Done—or returns with a precise next action.
Telegram is an interface, not the database
Chat is excellent for quick capture and notification. It is poor at answering durable questions such as “what is blocked?”, “who owns this?”, or “which tasks passed review?” when those answers live only in message history.
Kangram connects Telegram, web, and MCP operations to the same backend task records. The practical rule is simple: discuss freely in chat, but put commitments, ownership, state, and acceptance in the task.
Capture enough context without writing an essay
A useful Telegram capture contains:
Fix the onboarding redirect after Google login.
Outcome: new users land on board creation, not an empty dashboard.
Priority: Major. Owner: agent after Roman confirms expected route.Triage can add code links, screenshots, and acceptance checks later. The first message should make the outcome recognizable and prevent a vague “look into login” task.
Let the agent work from the board
Once connected through the Kangram MCP setup, an agent can query assigned or pending work, inspect a selected task, and apply permitted updates. A good start instruction is:
List tasks assigned to you on the product board. Do not start a task that is
already In Progress. For the selected task, summarize its outcome, constraints,
and acceptance checks before changing code or board state.This prevents the Telegram message from becoming a hidden prompt. The task is the reviewed handoff contract.
Keep human review explicit
Use a review state when an agent has produced an artifact but a human decision remains. The agent should attach or link the result, record validation, and name the decision needed.
For a reliable packet, use the async human-agent handoff template. For permissions and risky actions, apply the MCP security checklist.
Handle failure without splitting the backlog
If an MCP call fails, do not create a replacement task in chat. Keep the original task and record:
- the attempted operation;
- the exact error without secrets;
- whether any state changed;
- who can unblock it;
- the safe retry condition.
If Telegram is unavailable, the web or agent can still use the shared board. If the agent is unavailable, a teammate can continue from the same task. That surface independence is the point.
A small-team operating policy
- Humans approve membership, deletion, bulk changes, and external side effects.
- Agents may read approved boards and propose or create scoped work.
- One task has one current owner and one next action.
- Important decisions are copied into the task; secrets never are.
- Done means acceptance passed, not merely that a message was sent.
Start with one workflow
Choose one recurring handoff—bug triage, content release, customer follow-up, or weekly operations. Run it through Telegram, the board, and one agent for a week before expanding permissions or adding automations.
The MCP task-management guide explains the underlying model. The setup page provides the current Kangram connection instructions.