Descriptions:
AI Jason breaks down OpenAI Codex’s new `/goals` command and Hermes agent’s equivalent persist-goal feature — both designed to keep coding agents working autonomously on complex, long-running tasks without prematurely declaring success. The video explains the core problem these features solve: agents that stop too early on large or ambiguous projects, and how a persistent goal loop uses a separate LLM judgment call to determine whether work is genuinely complete rather than relying on a simple programmatic iteration counter.
The tutorial walks through how the goal loop works under the hood — from the initial goal submission through the LLM judgment prompt that decides whether to continue or stop — and compares Codex’s implementation (which asks the agent to self-report completion) with Hermes agent’s approach (which uses an external model as judge). Practical guidance covers how to write effective goal prompts: defining a clear definition of done, specifying what should not change, and providing verification mechanisms such as Playwright interactive tests to confirm output quality at each iteration.
Real-world use cases explored include overnight migration runs clocked at nine hours, large-scale refactoring, and optimization loops where an agent iterates until an evaluation score hits a target threshold. The video draws on learnings from Vincent, a maintainer of an open-source Claude toolchain, who ran the goal feature for three days across 30 rounds, underscoring the importance of aligning the agent on full context before launching any extended autonomous session.
📺 Source: AI Jason · Published May 09, 2026
🏷️ Format: Tutorial Demo
