How Hightouch built their long-running agent harness

3 months ago

Hightouch built a general-purpose marketing agent called Hightouch Agents, capable of planning campaigns, analyzing data, and automating marketing tasks.
The agent leverages the core Hightouch product for complete customer data context, including integrations with platforms like Facebook Ads and Hubspot.
Key challenges included managing long-running tasks, context limitations, and dynamic reasoning, which required innovative solutions like separating planning and execution phases.
Hightouch implemented dynamic plan updates, allowing the agent to adjust its strategy based on new data insights during execution.
To handle context limitations, the team developed techniques like buffering context to files and using dynamic subagents for isolated task execution.
Instead of relying on embeddings, Hightouch used a fan-out pattern with smaller, cheaper LLMs for tasks like analyzing ad creatives, which proved more effective and cost-efficient.
The agent harness includes features like system tool calls for internal thought management and frequent plan regurgitation to keep the model focused.
Hightouch's approach emphasizes practical 'context engineering' over theoretical advancements, focusing on real-world applicability and scalability.

Hasty Briefsbeta