What happened

xAI added a feature called Workflows to Grok Build, its command-line coding agent (you type instructions instead of clicking through a graphical app). You describe a large task in plain English, and Grok Build turns that request into a small script: a plan that breaks the work into phases, decides how many agents each phase needs, and defines how their results roll up into one final report. You never write that script yourself.

The official announcement walks through a pull request review as the example. Ask Grok Build to review a PR, and it plans four phases: gather context on the change, fan out specialist reviewers, have independent agents check each other’s findings, then write one ranked report. Each phase can run multiple agents at once, each starting with its own clean slate rather than sharing one long conversation.

A single workflow run can use up to 128 agents by default, or up to 1,024 for larger jobs. The work happens in the background, so your terminal stays free to do other things while it runs. Progress saves as the workflow goes, so pausing and resuming does not repeat finished work. Typing /workflows opens a live view of the run, phase by phase, including how many tokens (the units of text an AI model reads and generates, which is how usage against your plan is typically measured) each agent has used so far.

Workflows you want to keep are saved as files. Put one in .grok/workflows/ inside a project and it is shared with anyone who works in that repository. Put it in ~/.grok/workflows/, a folder tied to your own machine rather than any one project, and it follows you into every project you open.

Why it matters

Most coding agent sessions are a single back-and-forth conversation. That works for a focused edit, but it strains on a task with many independent parts, like checking every changed file in a large PR, triaging a backlog of open issues, or scanning a codebase for one class of bug. A single conversation has to hold all of that context at once, and a single reviewer, human or AI, misses things a large batch of work hides in.

Splitting the job into agents that each start fresh, then having other agents check their findings before anything reaches you, is a meaningfully different structure than one long chat. xAI calls this an “adversarially verified” pass: agents are explicitly asked to try to poke holes in earlier findings before those findings count. This kind of multi-agent orchestration, where one request becomes a coordinated batch of specialized workers instead of one continuous conversation, is becoming more common across coding agents generally, not just at xAI.

The tradeoff is that a workflow can spend up to 1,024 agents on one request. xAI’s announcement does not publish what that costs against a Grok Build plan, or whether the agent budget is metered separately from ordinary chat usage. Treat that as an open question until xAI documents it, not as a reason to assume it is free.

Who should care

This is most useful for a task you would otherwise avoid doing carefully because it is tedious: reviewing every file in a sprawling PR, working through a backlog of issues one by one, or auditing a codebase for a specific problem. If your typical task is a single small edit, a workflow adds planning overhead you do not need.

It also matters if you have used Grok Build before. xAI’s coding agent had a rough stretch earlier this month: it was caught uploading entire Git repositories to company storage without being asked, including files it never opened and old commits (saved snapshots of a project’s history) still holding secrets. xAI disabled that upload and later open-sourced the client harness so builders could inspect what the software actually does. A workflow runs many background agents against your code with less of your own attention on any single step, which raises the same question that prior incident did: what data leaves your machine, and when.

What builders should do next

If you already run Grok Build, try a workflow on a bounded task you can grade yourself, rather than something you have no way to check. Pick a pull request you already reviewed manually and remember well. Ask Grok Build to run a workflow review on it, then compare three things against your own earlier review: which real issues the workflow caught that you also caught, whether it flagged anything that was not actually a problem, and how many agents and how much time it used to get there. That tells you whether the adversarial verification step is catching real mistakes or just generating plausible-sounding notes.

To try it, install the command-line tool (CLI):

curl -fsSL https://x.ai/cli/install.sh | bash

That command downloads and runs xAI’s install script for the Grok Build command-line tool. Read a script before piping it into bash if you are not already comfortable trusting the source, since this pattern runs whatever the remote server sends with your user’s permissions.

Given Grok Build’s recent data-handling history, keep the same discipline that story taught: know what a workflow’s background agents can read and where their output goes before you point one at a repository holding real credentials.

End of article