The Data Layer: Creating a Virtual Market

GTM engineering aims to build a "perfect virtual copy" of the market. Because accounts are in constant flux—due to acquisitions, hiring, and internal sales activity—maintaining accurate records is a significant data engineering challenge.

• Waterfalling: No single data vendor provides a complete picture. The standard technique is "waterfalling," where you layer multiple providers to fill gaps (e.g., phone numbers). Teams must run automated evaluations (evals) against these providers to determine which data sources are most reliable for specific fields.
• Selective Refresh: Updating all data is cost-prohibitive. Engineers must prioritize which fields to refresh (e.g., employee count changes frequently, while headquarters location rarely does) and resolve entity conflicts between disparate third-party sources.

Orchestration and Graph-Based Logic

GTM teams often juggle 20–30 tools (CRMs, sequencers, dialers, etc.) that frequently disagree on the state of an account. Orchestration is the process of keeping these systems synchronized.

• The Sync Problem: Tools often sync independently (e.g., a CRM syncing to a sequencer). This creates race conditions where an action is triggered before data is available. Engineers must implement logic—such as waits and loops—to ensure data readiness.
• Graph Architecture: A robust orchestration layer should treat the system as a graph of general-purpose nodes (agents, tool calls, conditionals, code, and map-reduce fan-outs). This allows for complex event-driven workflows where a single trigger can fan out information to multiple systems and aggregate the results back.

Agentic Execution and Human Interfaces

Agents are now capable of handling long-running, unstructured tasks across a deal cycle of weeks or months.

• Persistent State: The most effective architecture involves one agent per account that maintains persistent state. These agents remain dormant until triggered by a heartbeat or specific event, at which point they ingest current context from the data layer.
• Execution Constraints: With cold email reply rates often below 1%, execution is high-stakes. Sending from a single domain risks reputation damage. Advanced teams use multiple domains for outreach and implement routing logic to ensure replies are funneled back to the correct sales rep.
• The Human Interface: The hardest problem in GTM engineering is the handoff between the agent and the human rep. Agents act as the reasoning and decision layer, but reps must remain in the loop. Designing systems where the human understands what the agent has done—and can override it—is critical for successful adoption.