A week with Sonnet 5.5: ask it for ideas and it builds the thing instead
every · x · 2026-09-29
After a week of testing Sonnet 5.5, every's Tyler Nishida found the model treats requests to think as permission to act:
- Asked for video ideas → it produced finished videos
- Asked what part of a visualization meant → it built another simulation instead of answering
- Asked to use one skill → it loaded several at high effort
The eagerness sometimes delivers more than expected, but also leaves you reviewing work you never requested before deciding what to make. The fix: state where the brief ends — say "Ideas only. Don't build anything yet," or ask it to explain existing work before changing it. Full write-up published by every.
More from coding & agent
- AlignOPSD fixes decision-timestamp mismatch in agent distillation, beating GRPO by 5.5-8.7% — Mingju Chen · 2026-09-29
- CompoWorld scales general agent training by composing environments from reusable services — AllSpark-Research · 2026-09-29
- Adaptive Consistency Graph lifts long-horizon agent success from 44.5% to 50.2% — Beihang · 2026-09-29
- ByteDance's TraceDance auto-builds agent behavior benchmarks from real deployment traces — ByteDance · 2026-09-29
- Databricks Tops All 4 NVIDIA SOL-ExecBench Kernel Tracks Using AI Agents for ~$70K — Yuchenj_UW · 2026-09-29
- ADHD is a superpower for juggling 10 AI agents, dinner, and shitposting at once — enggirlfriend · 2026-09-29