The Shift from Execution to Workflow Design

As AI models have matured, the primary bottleneck for enterprise adoption has shifted from the ability to execute tasks to the ability to understand and re-engineer complex, messy business processes. Most AI pilots fail to reach production because they are applied to broken, undocumented workflows. Forward Deployed Engineers (FDEs) are essential because they bridge the gap between high-level AI capabilities and the specific, idiosyncratic realities of an enterprise department (e.g., finance, logistics, or sales).

The Role of the AI-Augmented FDE

Varick Agents treats the FDE role as a high-touch, human-centered consultancy that avoids the "rip and replace" trap. Enterprises are often married to legacy systems like SAP or NetSuite and have zero appetite for migration. Instead of replacing these systems, Varick builds agents that sit on top of them. To scale this without exponential headcount, they developed internal tooling for their FDEs:

• Engagement Agent: Acts as an assistant that synthesizes raw data (Slack threads, PowerPoint slides, "granola notes") into a coherent workflow spec. It handles entity resolution (e.g., identifying that "Mike" in one system is the same as "Michael" in another) and queries process ownership.
• Workflow Agent: Embedded within their platform, this tool guides the FDE during construction, flagging missing edge cases or potential DAG (Directed Acyclic Graph) violations.
• Autonomous Assistant (In Development): Designed to handle routine maintenance—such as changing report routing or email recipients—based on client requests, freeing the FDE to focus on high-value strategy and client interviews.

Technical Implementation and Context Extraction

Frontier models often struggle with enterprise context because they are overly verbose and lack the nuance of a consultant who knows which details to prioritize. Varick addresses this through:

• Dependency Graphs: They represent business processes as dependency-driven graphs within a Postgres-backed source of truth, as enterprise workflows are often linear but prone to cycles.
• Post-training for Clarity: They post-train open-source models (e.g., Qwen/Kimmy) to balance detailed analysis with the brevity required for effective business operations.
• RL-based Traversal: They use Reinforcement Learning (RL) environments to train custom tools that traverse their knowledge graphs. These tools are specifically designed to handle enterprise-specific challenges, such as identifying redundant cycles or resolving entity ambiguity, which standard frontier models frequently fail to navigate reliably.