Grok Build adds autonomous /goal mode for long-running coding tasks with built-in verification

Grok Build now includes a new autonomous mode called /goal, which lets developers hand off a complex objective and step away as the agent plans, executes, and verifies its own work until completion. In a hands-off run, the agent builds a progress checklist, executes items sequentially, and runs verification tests before marking the task complete. Verification can occur by reviewing the produced code, inspecting runtime web pages, or executing tests, and the verification pass happens during the run rather than after it ends. This shift changes the typical workflow for terminal-based AI coding by enabling truly autonomous execution for bounded tasks.

The /goal mode uses a two-model pipeline: planning and instruction-following are handled by Composer 2.5, while Grok Build 0.1 handles code generation and execution. Composer 2.5 was added to the Grok Build CLI on June 1 to support long-running tasks. Grok Build 0.1 is xAI’s dedicated agentic coding model, designed for fast interaction with large codebases and multi-step tasks. The architectural choice opens questions about how independently the verification model operates from the generator, a topic developers will monitor as the feature matsures in production.

All code execution in Grok Build occurs locally on the developer’s machine; nothing in the codebase is transmitted to xAI’s servers during a session, a design choice highlighted for teams in regulated industries that require data locality and strong security.

Pricing and access remain a differentiator. The /goal mode requires a SuperGrok or X Premium Plus subscription, with tiers such as SuperGrok ($30 per month), SuperGrok Heavy ($300 per month), and X Premium Plus ($40 per month). At the base level, Grok Build is positioned as the lowest-priced full-featured terminal coding agent in its category.

Caveats exist. The SWE-Bench Verified score for xAI’s production Grok Build model has not been published. The benchmark figures cited come from an earlier model (70.8%), while Claude Code and Codex CLI report higher scores on SWE-Bench Verified. Elon Musk himself noted that production-scoped results would take longer to align with Claude’s performance, underscoring that long-horizon verification could shift which metric matters in practice.

FAQs

• What is the /goal mode and how does it work? It hands off a task to the agent, which plans, executes, and verifies until completion without requiring ongoing developer prompts. The agent creates a progress checklist and runs verification before finishing.
• What subscriptions are required? Access requires a SuperGrok or X Premium Plus subscription. Tiers include SuperGrok ($30/mo), SuperGrok Heavy ($300/mo), and X Premium Plus ($40/mo).
• How does verification work during long runs? Verification happens during execution and can involve reviewing code, inspecting runtime pages, or running tests to ensure the final deliverable works.
• Is execution local or sent to servers? All execution happens on the developer’s local machine; no code is transmitted to xAI servers during a session.
• How does Grok Build compare to Claude Code or Codex CLI? Claude Code and Codex CLI show strong published SWE-Bench Verified numbers, whereas Grok Build’s production-model score has not yet been published. The argument for /goal emphasizes long-horizon autonomous execution and built-in verification as a potential new priority metric.
• What about Arena Mode and availability? Arena Mode-allowing parallel agent runs against the same problem to select the best output-was confirmed in code traces but not yet live in the current beta; its release would affect how outputs are chosen in practice.

Source note: Coverage and details reflect the primary article describing /goal’s introduction and capabilities, with contextual references to the broader market landscape.