Descriptions:
Claravo, a product leader and AI builder, demonstrates how to use the /goal command in OpenAI Codex to run fully autonomous, multi-hour coding tasks — explaining both the mechanics of the feature and showing two real production use cases in action.
The /goal feature works differently from a standard chat prompt: instead of a turn-by-turn exchange, the agent sets a measurable success state and then loops autonomously — planning a step, executing it, verifying the result, and replanning — until the goal is met. Claravo explains the four lifecycle commands (/goal to set, pause, resume, and remove) and contrasts this with the REPL loop pattern others use in Claude Code. The first case study shows how Claravo used /goal to systematically work through every Sentry trace for an invalid edit-tool operation, categorize the root causes, implement fixes, and replay historical events — ultimately reaching zero remaining errors after several hours of unattended execution. The second is a live demo targeting Vercel API errors on a chat endpoint: the goal prompt instructs Codex to classify each error category, open a branch and PR for user-facing issues, and downgrade backend noise to warnings, then report back with a summary of all PRs and blockers.
The video is particularly useful for developers wondering how to move beyond prompt-and-wait workflows toward genuinely autonomous agents using GPT 5.5 in Codex.
📺 Source: How I AI · Published May 27, 2026
🏷️ Format: Tutorial Demo
