Workflow
AI Platform
AI Agents
Designing Multi-Agent
Workflow Configuration
Overview
At C3 AI, I led the end-to-end design of Agentic Workflows — a configuration experience that enables Reliability Engineers to create, customize, and automate multi-step AI workflows using natural language and a visual canvas.
This project focused on helping users define how AI should work, rather than just consume its outputs.
I designed the full user journey, from drafting a workflow with AI assistance, to configuring steps manually, to testing, publishing, and versioning it safely across production environments.
Role
Lead Product Designer ( 0→1 )
Duration
8 Weeks
Team
3 UX Designers, 3 PM, 8 Engineers, 2 Data Scientists
Discover the needs
When I started working with Reliability Engineers, I kept hearing the same thing:
Every time an alert comes in, I do the same steps — open the sensor data, check patterns, summarize, and create a work order.
Everyone had their own way of doing it, but the logic was always the same. The problem was scale — hundreds of assets, thousands of alerts, and endless repetition.
That’s when we asked:
What if engineers could teach the system their process once, and let it handle the next hundred times automatically?
The Problem
Reliability Engineers often deal with:
Too many alerts – each requires similar steps to investigate.
Inconsistent investigations – depending on who handles them.
Manual handoffs – copying findings or creating work orders manually.
Our existing GenAI agent could already handle one-off questions like:
“Which compressors are at risk?” or “What should I check next?”
But it couldn’t remember how a user wanted things done. Each investigation was a one-time conversation. Engineers needed a way to define their own process once, in a reusable, configurable way that reflected how they actually worked.
Reliability Engineers
Investigate asset health alerts daily.
Subject Matter Experts
Define investigation best practices.
Data Scientist
Define agents and tools that generate the best performance.
Designing for configuration,
not just creation
We didn’t want Agentic Workflows to feel like “coding in disguise.”
So the goal was simple:
Let engineers describe their process in plain language, and then give them the tools to see, edit, and control how the system performs each step.
Solutions
Create a Workflow - From Scratch or From a Template
When users open Agentic Workflows, they no longer face a blank screen or need to write code.
They can start from a prebuilt template that reflects common Reliability tasks, such as Alert Triage or Maintenance Analysis, or create a new workflow from scratch.
Seeing the Process Come to Life
The workflow opens on a visual canvas, where every step becomes a node.
Users can build in two ways:
Manual Mode — drag and drop nodes like Fetch Sensor Data or Run Anomaly Detection.
AI-Assisted Mode — let the LLM suggest new steps or generate entire sequences based on intent.
Once the structure is in place, engineers can fine-tune each step: adjust prompts, set thresholds, add input data, or swap tools. This balance between structure and flexibility gives users confidence — they’re not just using a fixed system, they’re shaping it.
Test Workflow with real world example
Engineers don’t need to get everything perfect on the first try. They can run a draft workflow in test mode, observe how the AI executes each step, and view generated summaries, anomalies detected, or any failed steps.
Instead of debugging code, they debug visually, tweaking a prompt or changing a tool right on the canvas. This gives users a sense of control and instant feedback, turning experimentation into a low-risk, learn-by-doing process.
Publish Workflow
Once a workflow works as intended, engineers can publish and schedule it to run automatically.
What used to be a manual routine now runs in the background. The same workflow can be reused across assets or shared with teammates, creating a library of best practices that captures the organization’s expertise.