Grok Bot by SpaceXAI
TL;DR Highlight
Grok Bot gives agents their own computer to log into apps and finish jobs end to end
Who Should Read
Developers who want to hand off repetitive work to agents but rely on internal tools with no API or MCP
Core Mechanics
- Each Bot has its own cloud computer
- Bots sign into apps, tools, and websites from their own computer.
- They reportedly work with platforms that have no clean API or MCP.
- Jobs keep running while you are away.
- You hand off work by messaging
- You message a Bot from phone or desktop like texting a colleague.
- The post contrasts this with tools that require building workflows first.
- Bots come back only when approval is needed.
- Showing a task turns it into a routine
- Ask the Bot to follow along the next time you do a job.
- It saves the steps as a routine and applies your corrections.
- Next time it runs the routine on its own.
- Multiple Bots coordinate with each other
- Bots message each other and share context in threads.
- In group chats they pass work and assign ownership.
Community Reactions
How to Apply
- Pick one repetitive no-API task
- Choose a recurring task your team does only through a UI, with no API or MCP.
- Write its input, steps, and approval points in five lines.
- Use this as the baseline when you demo or instruct an agent.
- Define approval points first
- List the steps where a human must confirm before the agent continues.
- Include hard-to-undo actions like sending email, payments, or closing tickets.
- Require drafts only for those until a human approves.
- Split Bot roles by lane
- Separate owners for lanes like inbox, expenses, and bug fixes.
- Give each the minimum accounts and permissions.
- Start the beta test with a single lane.
Try It
# Paste into Claude or ChatGPT (simplified version of the post's idea: writing a delegation spec for an agent)
I want to hand a repetitive task to an AI agent. Turn the task below into a routine spec the agent can complete end to end, with human-approval points separated out.
Task description: {{describe the task in the order a human does it}}
Tools used: {{CRM, Gmail, etc.}}
Output format:
1. Numbered step-by-step procedure
2. Input and output of each step
3. Hard-to-undo actions and where approval is required
4. Conditions under which the agent should ask a human
Next step: pull out only item 3 and agree on approval rules with your team.Terminology
Related Papers
Same Simulation, Different Evidence: Where Retrieved Tool Output Costs LLM Agents Accuracy
Saving long tool output to files cut GPT-5.1 accuracy by 11.3 points; naming the authoritative file restored it.
Fake News, Real Crashes: Information Injection Attacks on LLM Financial Trading Agents
Fake headlines hijacked LLM trading agents 62% of the time versus 14% for prompt injection.
Harness Engineering 101: The LLM is the brain, the harness is the body
An agent is a loop managing a JSON array; 16 chapters build one from scratch
Harness2 by Google Research – Recursive Agent Harnessing
Recursively refining the harness with fixed model weights raised benchmark scores by up to 11.8 points.
Ousted Savannah police officer gave ICE agent husband access to Flock system
A police officer shared her Flock login with her ICE-agent husband; an audit found 983 searches and she was fired
Tmux and Claude Code: Sessions, Send-Keys, and the Ceiling
Running Claude Code agents in tmux works well, but limits appear around four agents.