Taming AI-Assisted Code with Deterministic Workflows (obeli.sk)

Obelisk introduces deterministic workflow execution for AI-assisted coding by running workflows inside a WASM Component Model runtime and recording every child execution, parameter, return value and stateful action in an immutable execution log. That determinism enables exact replay, time-travel–style debugging (backtraces captured at each step), and an auditable trail of what the system did and why. Activities run in isolated VM-like sandboxes with scoped secrets, strict resource limits, and granular I/O permissions (disk and Process APIs disabled by default). Panics, OOMs and timeouts are converted into explicit errors the parent workflow must handle, forcing cleanup and compensation logic instead of silent failures. The practical upside for AI/ML teams is safer, reproducible, and debuggable human+LLM workflows—taming “vibe coding” by narrowing the blast radius of generated code and making behavior inspectable and repeatable. A Fly.io orchestrator case study shows a schema-first WIT interface, Rust bindings via wit-bindgen, and iterative LLM (Gemini 2.5 Pro) assistance producing HTTP client skeletons and serialization helpers; when a TOML serialization error occurred, the immutable logs and HTTP traces allowed rapid diagnosis and repair. For production AI systems, deterministic workflow engines like Obelisk promise stronger security, compliance, and operational confidence when combining LLMs with real-world side effects.