Scaling Forward Deployed Engineering with Internal AI Agents

The Forward Deployed Bottleneck

Most generative AI projects fail to reach production because they are applied to broken processes without understanding the specific, messy reality of an enterprise. Varick Agents addresses this by deploying engineers to map the actual, non-documented workflows of a department, such as the specific reconciliations between purchase orders and invoices. The goal is to re-engineer these processes around AI, automating the majority of steps while keeping humans in the loop for high-risk or non-standard tasks.

Engineering the Workflow Agent

To scale the productivity of their forward deployed engineers, Varick built an internal toolset called the FD Agent. This system functions in three stages:

• Engagement Agent: Acts as an assistant that synthesizes raw documentation, Slack threads, and notes to answer queries about process ownership and entity resolution (e.g., identifying that two different spellings of a name refer to the same person).
• Workflow Agent: Embedded within the Varick OS platform, this agent monitors the workflow construction process to ensure it adheres to the company's dependency graph and flags missing edge cases.
• Autonomous Assistant: A developing stage designed to handle minor configuration changes (e.g., routing updates) autonomously, allowing engineers to focus on high-value client strategy.

Technical Implementation

Varick avoids migrating clients off legacy systems like SAP or NetSuite, instead building agents that live on top of these systems of record. They represent business logic using a dependency graph and utilize custom-trained models to overcome the verbosity and lack of domain focus found in frontier models. They train these models in an RL environment with custom tools specifically designed to traverse their knowledge graph, identify redundancy cycles, and resolve entity conflicts.