Don't Vibe Code, Do Skele-Code: Interactive No-Code Notebooks for Subject Matter Experts to Build Lower-Cost Agentic Workflows
New research introduces a 'code-first, agent-assisted' paradigm to slash token costs in agentic workflow creation.
A new research paper titled "Don't Vibe Code, Do Skele-Code" proposes a novel interface for building AI agent workflows aimed at non-technical users. Developed by Sriram Gopalakrishnan, Skele-Code combines a natural-language and graph-based interface within an interactive notebook, allowing subject matter experts to incrementally construct workflows. Its core innovation is a "code-first, agent-assisted" paradigm where the AI agent is used solely for code generation and error recovery, not for the continuous orchestration or execution of tasks. This architectural shift is designed to drastically reduce the token consumption and associated costs of running complex, multi-step automated processes.
The system emphasizes modularity and reusability. Workflows built in Skele-Code are converted into structured code skeletons, making them easily extensible and shareable. Furthermore, a completed workflow can be packaged as a standalone "skill" that can be invoked by other agents or integrated as a step within a larger, more complex workflow. This approach directly challenges the prevailing "vibe coding" or fully agent-driven methods, prioritizing control, cost-efficiency, and the creation of durable, inspectable automation assets over fully autonomous but opaque and expensive agentic systems.
- Uses a 'code-first, agent-assisted' model where AI agents handle only generation & error recovery, cutting orchestration costs.
- Provides a natural-language and graph-based notebook interface for incremental, interactive workflow building by subject matter experts.
- Generates modular, shareable code skeletons that can be reused as agent skills or steps in other workflows.
Why It Matters
It enables cost-effective, scalable automation by letting domain experts build durable AI workflows without deep coding knowledge.