Once your agent is paired and running, controlling it is as simple as sending a Telegram message. Type a prompt in natural language and the agent starts working immediately. Output streams back to Telegram in real time.
Basic usage
Open Telegram, find the RemoteAgent bot, and type any prompt:
List all the files in this project and summarize what each one does
Fix the TypeScript error in src/api/users.ts
Add input validation to the POST /api/products endpoint using Zod
Write unit tests for the authenticate() function in lib/auth.ts
Refactor the database queries in services/orders.ts to use transactions
Streaming output
Output is streamed in real time from your agent to Telegram. For long responses, messages are split automatically at Telegram’s 4096-character limit. You may see multiple messages appear in sequence for a single command.
Streaming means you see results as they are produced — not all at once at
the end. For a typical “fix this bug” task, you will often see the AI’s
reasoning, the file it is editing, and the final diff, all appearing
progressively within a few seconds.
Aborting a running task
If you want to stop the agent mid-task — for example, if it is taking too long or going in the wrong direction — send:
The agent receives the abort signal via the Redis channel and terminates the current runner process. You receive a confirmation message. The agent then returns to the idle state and is ready for the next command.
Aborting stops the AI mid-execution. If the runner was in the middle of
writing a file, the file may be in a partial state. Check your working
directory after aborting if you used a runner with file-write capability
(claude-sdk, claude-code, aider).
Continuing a task
If the AI produced a partial result and you want it to continue from where it left off, send:
This is useful when the output was cut off or the AI indicated it needed more steps to finish.
Examples by category
Code review
Review the changes in the last git commit and point out any issues
Check auth.ts for security vulnerabilities
Bug fixing
I'm getting a 500 error on POST /api/checkout — fix it
The tests in __tests__/cart.test.ts are failing — investigate and fix
Feature development
Add a rate limiter to the /api/telegram route — max 10 requests per minute per user
Create a new API endpoint GET /api/stats that returns total session count for the current user
Explanation
Explain what the realtime.ts file does and how it connects to Supabase
What does the pairing flow do step by step?
Git
Show me the diff since yesterday
Commit all staged changes with a meaningful commit message
Message length and splitting
Telegram has a hard limit of 4096 characters per message. RemoteAgent handles this automatically:
- Output is split into consecutive messages at the character boundary
- Code blocks are preserved — splits never happen in the middle of a code fence when possible
- Very long diffs or file listings may produce several messages in sequence
There is no limit on how long a task can take or how much output it can produce. The relay will keep forwarding chunks as long as the runner is producing them.
When the agent is busy
If you send a command while the agent is still processing a previous one, the new command is queued. You will receive a message like:
Agent is busy — command queued. Send "abort" to cancel the current task.
