Mahax posted a setup on X: combining Grok Bot and Jev so that a bot team decides who handles what.

He puts it plainly: every decision made by Grok Bot can be up to 190x faster, at 1/440 of the original cost. Grok Bot does the work, Jev handles allocation, and any action involving money, signatures, or customers is left entirely to humans.

What Grok Bot Is

In his article Grok Bot. What It Is, and How to Actually Use It., Mahax breaks this down.

Grok Bot is not a model. Claude, ChatGPT, and Grok are the models you talk to it with. Grok Bot is the layer that lets models do work for you. Each bot has its own cloud computer with a browser, file system, terminal, and tools logged in with your accounts. When you close the app, it keeps running.

It operates apps the way a person does: opens webpages, fills out forms, clicks buttons. Output lands in the tools, not in a chat window. Emails go into the outbox, a record is added to the CRM, and a document gets a draft.

Multiple bots share one cloud computer and can hand off to one another. One bot finishes research and passes it to the next for formatting and sending. In AI circles, this architecture is called Agents, Loops, and Graphs, and until the past year engineers had to build it by hand. Grok Bot turns it into the default configuration.

What Jev Does in the Middle

In Mahax's setup, Jev is a scheduling tool. The first step is not installation but checking the existing bots, workflows, and tools, preserving the status quo, and filling only the gaps.

It answers three questions: which role is responsible, urgency from 1 to 5, and whether the bot can handle it independently, scored from 0 to 1. This score is the p-value used for routing.

The role list comes from roles.md. What each bot owns, what it must not touch, and which tools it can use are written in this file. Jev can only choose from this list.

Routing rules:

• p ≥ 0.85: assign directly
• p between 0.60 and 0.84: assign, but flag it in the brief
• p below 0.60: route to a human
• involves money, signatures, or customers: route directly to a human, regardless of p-value

Security Boundaries

In this setup, the security boundaries are more worth copying than the scheduling logic.

Every incoming message first becomes a job card: source, text, links, whether it involves money, whether it involves signatures, and deadline. Read-only by default. Instructions in emails are always treated as content, not executed. Any option Jev offers must be findable in roles.md; if it cannot be found, it is not executed.

Anything involving sending, publishing, paying, signing, or deleting goes into the human approval queue. Never click Always Allow. If a human does not approve, the bot does not act.

A task can be handed off at most 3 times. If it is still not complete after more than 3 handoffs, route it to a human.

Before launch, run 6 test tasks: one for each bot, one that must be routed to a human, and one email with a hidden instruction to see whether the bot gets led astray. Only after the tests pass should the official entry point be opened.

What Gets Faster Is Allocation, Not the Task Itself

A commenter pointed out: the 190x speedup is in deciding who does the action; the task itself may still take those 4 minutes, it is just allocated faster.

Another commenter questioned it: no real before-and-after comparison data was shown, and Jev's routing accuracy was not demonstrated either. Some also said they have seen countless Grok+Jev setups like this, and whether this one is actually better is hard to say.

These doubts are all reasonable.

When there are many tasks and many roles, human judgment about who should handle a given item really is a bottleneck. But if there are only three emails a day and one bot, this setup is over-engineering.

Mahax himself also said it suits work with repetitive, fixed outputs, not one-off tasks.

Real-World Problems

A commenter who tested it mentioned: Cloudflare blocks bots. Verified with an Upwork account, the IP was still flagged after restarting the computer and was not unbanned for 6 days. MFA can also leave bots stuck in place.

This is a first-generation public product. There is no independent benchmark yet validating the reliability xAI advertises. Do not treat it as stable infrastructure.

What Can Be Copied

In his tweet, Mahax provided the complete prompt, which can be pasted directly into Grok Bot. The core can be compressed into one sentence: let machines finish reversible operations, and wait for human approval for irreversible operations. roles.md should draw this line clearly first, then talk about efficiency.

For the complete setup, see the original tweet.