On June 22, 2026, xAI announced /goal, a new mode inside Grok Build, its terminal coding agent. Instead of stepping through each file edit in chat, you give the agent a single objective and it keeps working until the task is finished and verified. The ship lands the same week OpenAI expanded goal-style autonomy in Codex and GitHub pushed cloud agents to general availability, signaling that every major lab now treats "hand off and walk away" as a first-class CLI workflow.
One Line In, Checklist Out
Invoke /goal with a plain-language objective and Grok Build plans an approach, breaks work into a live progress checklist, and starts executing. You can still send additional instructions while it runs, but the default loop is autonomous: the agent owns planning, implementation, and verification rather than waiting for approval after every tool call.
xAI's example is straightforward:
/goal Migrate the auth module to the new APIWhen the run completes, the progress panel flips to Complete with every checklist item checked. That visual state matters for teams evaluating whether an agent actually finished a refactor or simply stopped generating plausible-looking diffs.
Built-In Verification, Not Just Code Generation
The distinguishing detail is what happens at the end of each step. xAI says /goal continues until work is completed and verified, not merely attempted. Verification can mean reviewing generated code, inspecting rendered webpages for front-end changes, or executing scripts to confirm behavior. An agent that edits files without testing its own output can report success before the change works; baking verification into the run closes that gap.
Reports from the launch describe subagents that implement and verify in rounds, sustaining momentum across longer tasks without the user re-prompting after every failure. That architecture mirrors what other vendors call judge loops or critic agents, but here it ships as a single slash command inside an existing CLI rather than a separate product SKU.
Steering Commands
Long-running autonomy still needs human override. Grok Build exposes four control commands inside an active /goal session:
/goal status— open the live progress panel/goal pause— stop work but keep the goal loaded/goal resume— continue from where it paused/goal clear— discard the current goal entirely
That split between pause and clear is practical: you can interrupt a run to inspect diffs or run your own tests, then resume without re-describing the objective from scratch.
Access and Installation
/goal is available now in Grok Build. xAI publishes a one-line installer for PowerShell and WSL environments; after install you sign in with your xAI account and invoke /goal from the terminal session. Access requires a SuperGrok or X Premium Plus subscription, so this is not a free-tier experiment like some open-weight CLI agents.
Grok Build itself reads your local codebase, runs shell commands, and supports MCP servers out of the box. /goal sits on top of that harness rather than replacing it, which means existing plugin and MCP configuration should carry over into goal-driven runs.
How It Compares to Codex Goal Mode
OpenAI shipped a similar concept earlier in June with Codex goal mode and Appshots, oriented around persistent objectives inside Codex sessions and visual validation of UI work. xAI's framing is more explicitly terminal-native: one slash command, a checklist panel, and verification through code review, page inspection, or script execution.
Neither approach removes the need for human review before merge. Both assume you trust the agent with an isolated branch or local tree long enough to reach a verifiable stopping condition. The difference is ecosystem: Codex ties into OpenAI's desktop app and GitHub-adjacent workflows; Grok Build targets developers already paying for SuperGrok who want autonomy without leaving the CLI.
What Web Developers Should Watch
If you evaluate Grok Build for production repos, treat /goal as a delegation mode with explicit exit criteria, not unattended production access. Start on branches, define verification you agree with (tests passing, specific URLs loading, lint clean), and use /goal status to audit checklist items before you merge.
The broader pattern is clear: interactive chat was the first phase of coding agents; autonomous goal completion with visible progress and built-in verification is the second. xAI's entry confirms that even vendors late to the terminal-agent race now ship long-horizon modes as table stakes.