Solo Dev Shares Honest Data: 163 Fugu Ultra Calls in Production

Future-Cook-6365 · reddit · 2026-08-13

A solo developer shares honest data and engineering reflections after 5 weeks and 163 API calls using the Fugu Ultra model in a production pipeline. The model is wired in as an on-demand "second brain" for heavy reasoning, multi-file review, and adversarial verification, following the core architecture rule: "Fugu proposes, primary agent verifies and authors."

Key Data & Lessons:

Lack of Verdicts: 92% of calls lacked a final accept/revise/reject verdict, making them hard to measure. The dev stresses building evaluation columns into logging systems from day one.
Manual Review (39 calls): 22 clearly useful, 11 stated known facts, 6 were noise.
Final Verdicts (13 calls): 3 accepted as-is, 9 accepted after revision, 1 rejected.

Core Engineering Patterns:

Fail-open client: Advisory models should never be load-bearing; the pipeline proceeds if the endpoint is down.
Route the HARD stuff: Noise grades almost entirely came from sending tasks a single model could handle.
Revision is a feature: Consensus output is a strong draft, not a final answer.

Original post →

Solo Dev Shares Honest Data: 163 Fugu Ultra Calls in Production

More from coding & agent