Opus 5 works well in simple coding, but breaks an autonomous Compound Engineering flow
danshipper · x · 2026-07-25
Kieran Klaassen says Claude Opus 5 is both impressive and annoying in an autonomous coding setup.
He reports that it works reasonably well at medium effort with minimal prompting, but it kept handing control back to the user inside Compound Engineering, even though the flow is meant to be autonomous. Because of that, he started rewriting Compound Engineering and is also building a skill to retune the harness/plugin. He says he’ll share more findings in the coming days.
More from coding & agent
- Internal CS research agent is already running with GPT-5.6 and Claude model choices — BenBajarin · 2026-07-25
- LiteParse 2.8.0 drops ImageMagick and speeds up image-to-PDF conversion up to 7.2× — llama_index · 2026-07-25
- Claude Opus 5 flips between 1/3 and 2/5 before finally settling on an answer — Sauers_ · 2026-07-25
- Opus 5 shines in async long-horizon work, with four agent patterns for real systems — RLanceMartin · 2026-07-25
- Why Software Factories Fail: Deep Dive into AI Coding Production Pitfalls — zeeg · 2026-07-25
- FrontierCode shows Claude Opus 5 peaking at medium effort, not max compute — philhchen · 2026-07-25