Descriptions:
Michael Automates tests Claude Fable 5 on a specific agentic coding challenge: autonomously generating, varying, and backtesting Pine Script trading strategies using a custom backtesting engine designed to replicate TradingView results. The 36-minute unattended run serves as a stress test of Fable 5’s extended agentic coding abilities.
The results are specific and cross-validated. Fable 5 identifies two new strategy variants — V1.6 and V1.7 of a momentum lock strategy — that outperform the previous champion on a risk-reward basis. The new top performer posts backtested returns of 21,656% profit with 30.1% maximum drawdown on Solana, and 6,136% profit with 38.2% drawdown on Fetch.ai. Michael manually verifies these numbers against TradingView’s own charting tools and confirms the figures match, giving the backtesting engine’s output meaningful credibility.
Beyond the headline results, the video candidly addresses the token cost involved in running Fable 5 at high effort levels, and notes the model’s 80.3% agentic coding benchmark score (versus Opus 4.8’s 69.2%) seems to translate to genuine real-world task performance — at least for well-specified, self-contained coding problems. For developers building AI-assisted quantitative or algorithmic workflows, this is a practical demonstration of what Fable 5 can and cannot do when left to run independently on a complex multi-step task.
📺 Source: Michael Automates · Published June 10, 2026
🏷️ Format: Hands On Build
