Sparqlite: What 12 Hours of Ultracode Agent Orchestration Reveals About Vibe Coding Limits

Описание и применение

Эксперимент показывает, что при использовании режима Ultracode от Claude Code можно быстро создать крупный код, но без наблюдаемости и аудита человек не сможет понять и поддерживать его.

Задача

A human can delegate a large coding task to an AI orchestrator, but after the AI completes the task the human cannot feasibly review, understand, or maintain the output.

Результат

The experiment demonstrates that while Ultracode orchestration can produce large, functional codebases quickly, it leaves the human unable to audit or maintain the output, exposing risks such as semantic drift, memory unsafety, and dependency rot.

Рабочий процесс

• 1. Prompt Claude Code to write a SQLite extension in Rust that implements SPARQL 1.1.
• 2. The Ultracode orchestrator partitions the task into subagents, generating JavaScript that defines a network of subagents.
• 3. Subagents produce multiple Rust crates: parser, planner, executor, storage, and FFI.
• 4. Subagents generate tests that pass 636 of 639 W3C SPARQL evaluation tests.
• 5. Subagents generate README files and a tutorial with example data.
• 6. The human reviews the output and finds it infeasible to audit the code.
• 7. The human identifies the lack of observability trails (decision logs, dependency traces, test coverage heatmaps, diff-to-spec alignment, performance profiles).

Доказательства

Simon Willison pointed Claude Code at a single prompt and walked away. Twelve hours later, he had a full SPARQL 1.1 implementation in Rust that runs inside SQLite, complete with multiple crates, READMEs, and a tutorial.

Источник

The agent ran for 12 hours. A human review would take weeks.

Источник

None of these exist in Claude Code's Ultracode mode today. The agent produces code and documentation. It does not produce a review trail.

Источник