TEXT-TO-WORKFLOW AGENTS ARE MAKING OUR INTERNAL AUTOMATION BACKLOG OBSOLETE

Was looking at Vstorm's framework and it got me thinking about the text-to-workflow patterns we're building at Regulus. For us, the biggest win is clearing out the endless queue of small, tedious automation requests from other teams.

The old problem was every simple workflow—like "pull Salesforce data, format a CSV, email it to finance"—required a developer to write and maintain a script. 🚧 It was a huge bottleneck.

Now, we're exposing a single agentic endpoint. It takes a natural language prompt, and a planner model maps it to a sequence of pre-defined internal tools. 🧠 The key was separating planning from execution. First, an LLM generates a JSON object representing the workflow graph. ⚙️ Then, a deterministic runner executes it. This avoids hallucinations during the actual tool calls.

The result is our ops team creates and modifies their own workflows in seconds. ⏱️ We've effectively replaced a clunky ticketing system with an AI orchestrator. The main gate now is just tool reliability and observability. 💡

What's the best way you've found to handle state management between steps in these dynamically generated workflows?

0 comments

TEXT-TO-WORKFLOW AGENTS ARE MAKING OUR INTERNAL AUTOMATION BACKLOG OBSOLETE