The NxtFruit blog
Notes from
the build.
Insights on AI development, technology choices and business outcomes, from the engineers who do the work.
AI Development
Sleeper Agent LLMs and the Untrusted Text That Shapes AI Plans
A review of sleeper agent LLMs led us to a narrower risk in our AI development platform: untrusted discovery answers can shape an early scope draft. Here is what the review supports and what we recommend changing.
· 6 min read
AI Development
Claude Code Skills From Real Engineers: The One Idea We Adopted for Bug Triage
A famous engineer's open-sourced skills directory, audited idea-by-idea: what we adopted (evidence-gated triage confidence), what we rejected, and the refutation method that decides.
· 5 min read
AI Security
Claude Code’s Auto Mode Can Be Broken — Here’s the Gap It Exposed in Our Pipeline
Security researcher Johann Rehberger broke Claude Code Opus 5's Auto Mode with a "confused-environment" attack — and when we checked the same mitigations against our own AI development platform, we found a live gap of our own to close.
· 7 min read
AI Development
We Reviewed Claude Plugins Against Our Codegen Pipeline — and Kept What We Had
Anthropic's Claude plugins bundle tools and skills for one-click install. We compared that model against our production AI development platform and found we already solve the same problem—with stronger trust and reproducibility gates.
· 7 min read
AI Development
Utopia Hit +3344 GitHub Stars. Our RAG Stack Already Had the Idea That Matters.
GitHub trending project utopia sells bitemporal facts as the fix for stale RAG. We compared it to our production AI development platform and found provenance and staleness already engineered—without the ontology weight.
· 7 min read
AI Development
Two AI Models That Can’t Hallucinate — and Why We Adopted Neither
Two "System One" decision models launched weeks apart — one proprietary, one Apache 2.0. They return typed values instead of text, so malformed output is structurally impossible. We reviewed both against our own AI development platform and adopted neither, and the reason had nothing to do with speed.
· 7 min read
AI Engineering
The Claude Code Source Leak: 3 Fixes It Found in Our Own Agent
A deobfuscated read of Claude Code's shipped bundle hit 1,376 points on Hacker News. We reviewed it against our own AI development platform — and the three changes that survived the review were all fixes to our code, not features copied from theirs.
· 8 min read
AI Engineering
AI Coding Agent Maintenance Costs: The Number We Weren’t Reporting
A Hacker News essay argued that an AI coding agent has to reduce your maintenance costs, not just your typing. We reviewed it against our own AI development platform and adopted one change — plus found a dependency gap we were manufacturing at provisioning time.
· 7 min read
AI Engineering
LLM Coding Agents Can’t Copy-Paste — We Found 49 Missing Lines in Our Own Repo
A widely-shared essay argues that LLM coding agents can't copy-paste — every "move" is really a delete-and-retype. We reviewed that claim against our own AI development platform and found the gap real, with 49 dropped rule lines in our own history to prove it.
· 7 min read