Astra vs Fable 5.1 on real ML tasks: Astra grinds agentically, Fable writes better

returnity · reddit · 2026-09-06

A Reddit user ran the same ML text-processing and model-training workflow on Astra and Fable 5.1 (both at xhigh), with sharply different results:

Astra is the better agentic coder: stricter eval protocol (70/15/15, held-out validation for model selection), root-caused a gensim 4.4 compiled-kernel bug and fixed the environment via dependency downgrades (Fable just hid the stderr), rebuilt an isolated uv venv, SHA-256'd the corpus, emitted split manifests and run summaries, and even QA'd via headless-browser screenshots. It also used the author's pre-built notebook-reviewer and citation-checker subagents, one of which caught and fixed a real tokenization bug with a regression test.
Fable writes better and follows instructions better: more idiomatic, readable code that adhered to the author's coding-conventions.md; handled UTF-8 correctly while Astra confidently shipped a Windows-1252 decoding defect that garbled currency symbols throughout the HTML output. Fable's analysis report was far more insightful, ablating the preprocessing pipeline to expose an expensive no-op step and flagging a subtle ranking discrepancy.
Both improved F1/Accuracy by only 0.02-0.04 after human feedback, showing neither has mastered the full ML text pipeline yet.

Verdict: pick Astra to autonomously grind through broken environments with a forensic audit trail; pick Fable 5.1 for prose and readable code.

More from coding & agent