Tibo
@thsottiaux
/goal might be the most consequential thing we have shipped in codex The value of good instructions has never been higher.
185retweets4.9klikes
View on XOpenAI Ships Codex /goal to Automate Multi-Day Engineering Workflows
OpenAI released the
/goal command for Codex, a terminal-based agentic coding tool that autonomously executes multi-step tasks. This feature introduces a long-horizon mode where the AI enters an autonomous agent loop (an iterative cycle of reasoning and action) to achieve objectives without constant human intervention.This update follows plans to provide Codex's long-running workflow controls that operate for days on complex problems. By moving beyond single-turn interactions, OpenAI is positioning natural language instructions as the primary substrate for Karpathy's agentic engineering framework (building reliable autonomous systems), shifting the development bottleneck from writing code to defining success criteria.
You can now invoke /goal followed by an objective, such as implementing a backlog or refactoring a module while maintaining test coverage. The command is available in the Codex CLI, which recently added Codex's background computer use to manage the software development lifecycle across multiple applications.
Still wondering? A few quick answers below.
The /goal command is a feature for the OpenAI Codex CLI that enables autonomous, long-horizon coding. Instead of issuing single instructions, users define a high-level objective. Codex then enters an agent loop, performing multi-step actions like navigating codebases, writing code, and running tests independently until the task is complete.
When a user invokes the /goal command, Codex initiates an autonomous agent loop. It observes the environment, reasons about the next necessary step, and executes actions across multiple turns. This allows the AI to handle complex engineering tasks, such as implementing a full feature backlog, by self-correcting and iterating until it meets the success criteria.
Users can manage autonomous sessions using specific slash commands within the Codex interface. The /goal command followed by an objective starts a new task. To stop the agent while preserving its current progress and state, users can use /goal pause. The /goal resume command allows the agent to restart and continue working toward the original objective.
The /goal command is available to users of the OpenAI Codex CLI and the dedicated Codex desktop application. OpenAI has confirmed that Codex remains accessible to users on both ChatGPT Free and Plus subscription plans. This ensures that individual developers can utilize these autonomous agentic coding capabilities regardless of their specific paid tier.
Standard Codex prompts typically involve single-turn interactions where the AI responds to one specific instruction. In contrast, the /goal command activates a long-horizon mode. This allows the AI to manage its own execution over many steps and turns, making it suitable for complex engineering objectives that require sustained autonomy rather than isolated code generation.
