Hightouch recently launched its Hightouch Agents product, designed to automate complex marketing tasks. This tool can plan campaigns, analyze data, and generate reports, all while maintaining context from various marketing sources like Facebook Ads and HubSpot. Building such an advanced agent required the Hightouch engineering team to tackle a range of unique challenges, particularly around context management and long-running tasks. Traditional agent frameworks, often rigid and based on Directed Acyclic Graphs (DAGs), fell short for the open-ended nature of marketing tasks, which require flexibility and creativity. One key innovation from Hightouch was separating the planning phase from execution. Instead of diving straight into action, the agent first lays out a step-by-step plan for addressing a prompt. For example, when asked about churn indicators, the agent would outline a series of analytical steps before executing them. This approach allows for real-time updates to the plan based on new data, enabling the agent to adapt its strategy as it works. This dynamic planning feature, facilitated through specific tool calls, enhances the agent's reasoning capabilities. The challenges of long-context reasoning in LLMs presented significant obstacles, particularly regarding context windows and the models' tendency to settle for satisfactory answers too soon. Hightouch's method aims to overcome these limitations by prioritizing a model's ability to think critically and adjust its approach as new information emerges. This makes the Hightouch Agents not just a sequence of operations but a more intelligent system capable of complex reasoning in the marketing domain.