Descriptions:
This video from the All About AI channel walks through a working agentic trading system built around a ‘heartbeat’ architecture: a lightweight GPT-4o-mini sub-agent polls live position data from a WebSocket every 30 seconds, compresses it into structured JSON, and feeds that summary to a primary Codex GPT-4.5 agent that makes or adjusts trading decisions based on the current position and a stated profit goal.
The demo runs on Hyperliquid with real capital — a $50 margin, 10x leveraged short on the S&P 500. The creator shows how the primary agent, on receiving the goal of ‘$1 profit in 30 minutes,’ calculates that a 15-point downward move is needed, sets a maximum acceptable loss, and then monitors the position through successive heartbeat cycles. The sub-agent is configured as read-only, running on the faster and cheaper GPT-4-mini, while the main Codex agent handles reasoning and order management.
The architecture pattern — separating a fast, cheap monitoring sub-agent from a more capable reasoning agent — is shown as a practical solution to a real agentic challenge: how to maintain continuous situational awareness in a live system without burning tokens on expensive reasoning models for every data refresh. The setup uses a markdown instruction file to define both agents’ roles, and the video demonstrates the Codex `/goal` command for updating the agent’s objective mid-session.
📺 Source: All About AI · Published June 01, 2026
🏷️ Format: Hands On Build
