Sophie: The Five-Agent Pipeline That Builds Workflows from Natural Language

The workflow development bottleneck

Every supply chain has unique processes: specific carrier escalation protocols, custom compliance rules, industry-specific documentation requirements, facility-specific scheduling logic. Building automation for these processes traditionally requires a consulting engagement, custom development, QA, and months of implementation. The cost and timeline make automation inaccessible for all but the highest-volume workflows.

The five-agent pipeline

Sophie decomposes the workflow development process into five autonomous stages, each handled by a specialized agent.

Agent 1: Analyze

The Analyze Agent receives the natural language requirement ('When an asset is not assigned 2 hours before its appointment, alert the operations team and suggest a replacement') and decomposes it into structured components: triggers, conditions, data dependencies, actions, and escalation paths. It identifies which Digital Twins are needed, which Loft actions exist, and whether any capability gaps require new action development.

Agent 2: Plan

The Plan Agent takes the structured decomposition and generates the workflow architecture: which Loft actions to call in what sequence, what data to pass between steps, what branching logic to apply, and what error handling to include. The plan is presented to the user for review on a visual canvas showing the workflow from trigger to completion.

Agent 3: Execute

The Execute Agent generates the code: the Agent Operating Procedure (AOP) definition, the Temporal workflow registration, the action wiring, and the API contracts. The generated code follows FourKites' internal standards and integrates with the existing Loft infrastructure.

Agent 4: Test

The Test Agent validates the generated workflow against historical scenarios from the Graph. If the workflow is designed to escalate late shipments on a specific lane, the Test Agent replays historical shipment data for that lane and verifies that the workflow would have fired at the correct times, with the correct actions, and without false positives.

Agent 5: Catalog Update

The Catalog Agent registers the new workflow in the Loft Workflow Library, making it discoverable, reusable, and available for deployment to other customers with similar requirements. Every workflow Sophie builds becomes a component that accelerates the next build.

Why this matters

Sophie is agents building agents. The five-stage pipeline transforms workflow development from a months-long consulting engagement into a days-long configuration process. More importantly, every workflow Sophie builds is tested against real data, deployed with full decision tracing, and cataloged for reuse. The platform gets richer with every customer requirement.