How to Orchestrate AI Workflows

2025-04-30 · Source: Refactoring · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Intermediate, short

Summary

This article explores the orchestration of AI workflows, contrasting pure AI-driven approaches with traditional orchestration platforms like Kestra. It uses a real-world example of automating release aftermath tasks for Tolaria, such as fetching GitHub commits, matching tasks to Todoist, updating GitHub Issues and Canny, and generating release notes. While initial pure AI workflows are fast to ship, they suffer from slowness, expense, silent failures, dirty states, and lack of built-in retries or recovery. The piece advocates for a hybrid model, where AI agents serve as "scaffolding" for prototyping and handling "messy" tasks like classification or summarization. Deterministic elements, such as routing, retries, and scheduling, are best handled by robust orchestration platforms, which offer superior observability, reliability, and scalability. A four-stage workflow engineering journey is proposed, culminating in AI-driven composability where LLMs generate structured workflow definitions for stable orchestration layers.

Key takeaway

For AI Engineers or MLOps teams building AI-driven automations, recognize that pure agentic workflows, while fast to prototype, introduce significant reliability and observability challenges. You should strategically integrate robust orchestration platforms for deterministic tasks like scheduling, retries, and structured output validation. Evolve your workflows by extracting stable components into explicit code, reserving AI for complex judgment calls. This hybrid approach ensures scalable, maintainable, and debuggable systems, moving towards AI-driven workflow generation.

Key insights

Pure AI workflows are fast to prototype but lack reliability; combine AI with deterministic orchestration for robust systems.

Principles

• AI excels at messy tasks, not infrastructure.
• Orchestration provides reliability, observability, scalability.
• Use agents for prototyping, then harden with code.

Method

The workflow engineering journey progresses from agent-first prototyping to isolating deterministic parts, then structured orchestration, and finally AI-driven composability where LLMs generate workflow definitions.

In practice

• Extract deterministic steps into explicit code.
• Use AI for classification, summarization, judgment calls.
• Adopt code-first, API/CLI-first, open-source orchestration.

Topics

• AI Workflows
• Workflow Orchestration
• AI Agents
• Kestra
• MLOps
• Software Engineering

Best for: AI Engineer, MLOps Engineer, Software Engineer

Related on AIssential

• Build, Reuse, or Hybrid? How Orchestration Powers Agentic AI — Effective orchestration is crucial for integrating and coordinating agentic AI components, whether built, reused, or hybrid.
• Workflows for work that runs the business - Mistral AI — Workflows provides durable, observable, and human-in-the-loop orchestration for enterprise AI applications, moving them reliably to production.
• AI agents are entering their rebuild era as enterprises confront the reliability problem — Enterprise AI agent reliability in production hinges on robust workflow orchestration, state management, and recovery, not just LLM performance.
• Building durable Agents with Workflow DevKit & AI SDK - Peter Wielander, Vercel — Workflow DevKit simplifies building durable, observable, and resumable AI agents by abstracting complex orchestration patterns.
• Inside Multi-Agent Orchestration: State, Security, Scaling and Future — Multi-agent orchestration coordinates specialized AI agents to achieve complex shared goals, mirroring project management for AI workflows.
• Agents as scaffolding for recurring tasks. — Hybrid agent-code workflows enhance reliability for recurring tasks by leveraging agents for ambiguity and code for precision.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Refactoring.