A Durable, LLM-Aware Workflow Automation Platform Built on Serverless Step Functions

Abstract

Modern business automation increasingly mixes deterministic operations such as HTTP requests and message dispatch with non-deterministic generative steps powered by Large Language Models. Existing low-code automation tools, including Zapier, Make and n8n, treat LLM calls as opaque HTTP nodes and offer little support for the reliability, observability and credential safety that production LLM workflows demand. This paper presents Nodebase, a self-hosted workflow automation platform that treats LLMs as first-class nodes in a directed acyclic graph and delegates durable execution to a checkpointed serverless step substrate. Workflows are constructed in a browser-based visual editor and executed by an Inngest-backed function whose individual node steps are independently retried, memoized and streamed back to the editor through per-node real-time channels. We describe the system architecture, the workflow data model, the executor registry pattern that makes new node types a closed-form extension, and the prompt-templating layer that lets downstream nodes consume upstream outputs by name. A controlled evaluation on workflows of varying size shows that automated execution reduces end-to-end completion time by 12× to 26× relative to a human operator performing the same steps manually, with sub-15-second wall-clock latency for six-node workflows and a measured 96.4% success rate under simulated transient API failures. We discuss limitations, in particular the dependence on a hosted durability layer, and outline a path toward conditional, looping and agentic node types.

Introduction

Modern workflow automation is increasingly combining traditional integration tools with Large Language Models (LLMs), but existing systems struggle to handle them efficiently. iPaaS platforms like Zapier or n8n allow visual workflow building, but treat LLM calls as simple HTTP requests, leading to poor error handling, weak observability, and no structured support for retries or credentials. On the other hand, LLM-focused frameworks like LangChain and LangGraph offer powerful model orchestration but are code-first and lack user-friendly visual interfaces. Durable execution systems such as Temporal and Inngest solve reliability through checkpointing and retries, but they are infrastructure tools rather than complete end-user products.

To address this gap, the paper proposes Nodebase, a workflow automation platform that integrates all three paradigms. It provides a visual DAG editor for easy workflow creation, first-class typed LLM nodes with support for multiple providers, and a durable execution system where each node is independently checkpointed. This design enables reliable retries, better observability, and easier prompt/data handling across nodes. It also introduces an executor registry for extensibility and real-time per-node monitoring for easier debugging.

Nodebase is built using modern web technologies and evaluated on synthetic workloads, showing fast execution and high reliability even under failures. Overall, it aims to unify visual workflow design, LLM orchestration, and durable execution into a single, practical system that is easier to use and more reliable than existing tools.

Conclusion

We presented Nodebase, a workflow automation platform that treats LLM calls as first-class typed nodes, executes each node as a checkpointed step on a durable runtime, and exposes per-node status through realtime channels back to a browser-based visual editor. The system bridges three previously distinct categories of tooling: iPaaS visual editors, LLM orchestration libraries and durable execution engines. A controlled evaluation showed end-to-end speedups of 14–26× over manual execution on representative workflows and a 96.4% success rate under aggressive transient-failure injection. Several limitations of the current implementation deserve emphasis, since reviewers and prospective adopters will surface them. First, durable execution is delegated to a hosted Inngest tier, which makes the system dependent on an external service. The Inngest CLI provides a self-hosted development server, and a fully self-hosted Inngest deployment is on the project roadmap, but the production prototype is not yet detached from the hosted control plane. Second, the workflow model currently supports straight-line and branching DAGs but does not yet support conditional routing or loops; these are the principal expressivity limits compared to LangGraph. Third, the credential model is single-tenant: each user owns and references their own keys, with no organization-level sharing. Fourth, we have not yet conducted a user study to evaluate the editor’s usability against existing iPaaS tools. Several avenues of future research follow naturally: 1) Conditional and looping nodes. A small set of new node types would extend the expressiveness of the DAG without changing the executor contract: a conditional branch, a parallel split-and-join, and a bounded loop. 2) Sub-workflows and reuse. Allowing a node to invoke another workflow would enable users to factor large automations into composable units. 3) Agentic LLM nodes. An iterative tool-using LLM node, similar in spirit to LangGraph’s agent loop but expressed as a single node in the editor, would let workflow authors mix one-shot LLM calls with deliberative agents in the same canvas. 4) Self-hosted durability. Decoupling the prototype from the hosted Inngest control plane and documenting a fully on-premises deployment is essential for adoption in privacy-sensitive contexts. 5) Empirical user study. A controlled study comparing time-to-first-working-workflow on Nodebase against n8n and Lang-Graph would substantiate the qualitative claims made in this paper.

References

[1] Saxena, “n8n: An Open-Source Workflow Automation Platform for Enterprise Integration and AI-Driven Orchestration,” International Journal of Computer Applications, vol. 187, no. 63, Dec. 2025. [2] Flowmondo, “n8n vs Zapier vs Make: Which Automation Tool Is Right for You,” Online article, 2025. [3] WeMakeFuture, “iPaaS Comparison 2025: The 5 Best Tools,” Online article, 2025. [4] Vellum AI, “15 Best n8n Alternatives in 2026,” Online article, 2026. [5] LangChain Inc., “LangGraph: Agent Orchestration Framework for Reliable AI Agents,” Documentation, 2025. [6] Vercel, “AI SDK: The TypeScript Toolkit for AI Applications,” Documentation, https://ai-sdk.dev/, 2025. [7] xyflow, “React Flow: Node-Based UIs in React,” Documentation, https://reactflow.dev/, 2025. [8] D. Farrelly and T. Kohler, “The Principles of Durable Execution,” Inngest Engineering Blog, 2024. [9] Temporal Technologies, “The Definitive Guide to Durable Execution,” Documentation, 2024. [10] Apache Software Foundation, “Apache Airflow Documentation,” https://airflow.apache.org/, 2024. [11] Prefect Technologies, “Prefect Documentation,” https://docs.prefect.io/, 2024. [12] Amazon Web Services, “AWS Step Functions Developer Guide,” Documentation, 2024. [13] Vercel, “Next.js: The React Framework for the Web,” https://nextjs.org/, 2025. [14] Prisma, “Prisma ORM Documentation,” https://www.prisma.io/docs, 2025. [15] A. Johansson et al., “tRPC: Move Fast and Break Nothing,” https://trpc.io/, 2025. [16] Better Auth, “The Most Comprehensive Authentication Framework for TypeScript,” https://www.better-auth.com/, 2025. [17] Handlebars.js, “Minimal Templating on Steroids,” https://handlebarsjs.com/, 2024. [18] A. B. Kahn, “Topological Sorting of Large Networks,” Communications of the ACM, vol. 5, no. 11, pp. 558–562, 1962. [19] M. Harchol-Balter, Performance Modeling and Design of Computer Systems: Queueing Theory in Action, Cambridge University Press, 2013. [20] R. T. Fielding, “Architectural Styles and the Design of Network-based Software Architectures,” Ph.D. dissertation, University of California, Irvine, 2000. [21] A. Vaswani et al., “Attention Is All You Need,” Advances in Neural Information Processing Systems (NeurIPS), 2017. [22] Anthropic, “Claude API Documentation,” https://docs.anthropic.com/, 2025. [23] OpenAI, “OpenAI Platform Documentation,” https://platform.openai.com/docs, 2025. [24] Google, “Gemini API Documentation,” https://ai.google.dev/, 2025. [25] OWASP Foundation, “OWASP Top Ten,” https://owasp.org/www-project-top-ten/, 2024. [26] Sentry, “Application Monitoring and Error Tracking,” https://sentry.io/, 2025. [27] Stripe, “Webhooks API Reference,” https://stripe.com/docs/webhooks, 2025. [28] M. Jones and D. Hardt, “The OAuth 2.0 Authorization Framework: Bearer Token Usage,” RFC 6750, IETF, 2012. [29] S. Ko et al., “A Survey on Visual Programming for Data Science and AI,” ACM Computing Surveys, vol. 55, no. 9, 2023.

Copyright

Copyright © 2026 Aniket Kumar Mahato, Shristi Chaurasia, Abhishek Chauhan, Arundhati Walia, Yash Raj Nishad, Traymbakesh Mishra. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.