Hightouch built a general purpose marketing agent that handles complex, multi-step tasks like campaign planning and data analysis. Their engineering team solved problems with context management and long-running workflows that existing frameworks couldn't handle. The system separates model planning from execution for better results.

Traditional agent frameworks rely on rigid Directed Acyclic Graphs, which fail for open-ended marketing tasks. Hightouch's approach uses agentic delegation with dynamic plan updates as the agent learns from customer data. This lets the system adapt when encountering unexpected insights during execution.

The harness buffers context to files and fans out to smaller, cheaper models instead of embeddings. This addresses LLM limitations like context window constraints and default concise behavior. The result is one of the most advanced agent systems in production today, handling sparse queries across diverse customer bases.