How Apollo Leverages Deep Agents and LangSmith for GTM AI

As a Go-To-Market (GTM) platform, Apollo faced the challenge of complexity, requiring users to manually operate multiple modules as its product features expanded. To address this, they built the Apollo AI Assistant, which executes the entire loop from lead generation and data enrichment to outreach and measurement using only natural language commands.

Previously, Apollo used a Supervisor architecture based on LangGraph, but faced limitations in having to manually code and connect sub-agents for every new use case. Consequently, Apollo redesigned its architecture with Deep Agents to ensure model neutrality and support flexible execution.

Key features of the new Deep Agents architecture include:

• Dynamic Skill Selection: The agent autonomously determines and executes the necessary skills (such as lead generation, sequence creation, and analysis) based on the user's goal.
• Maximized Development Efficiency: By adopting a skill-based flat structure, the workload from initial development to launch was reduced by approximately 80-85%.
• Improved User Experience: By reducing unnecessary confirmation steps and improving response latency, it provides a more natural conversational interface.

Additionally, Apollo utilizes LangSmith to trace and debug the execution of its complex multi-agent system. This allows engineers to instantly identify which tools were called and in what order via specific thread IDs, rather than manually cross-referencing logs. Based on this, Apollo has established a six-layer evaluation framework called AI Watchtower to manage AI quality.