Goal Engineering: The Future of Designing AI Agents

According to LangChain Blog:

Background and Challenges

As AI agents transition from simple text generation and tool calling to executing full, open-ended tasks, the way we design their scope matters immensely. Pointing agents in the wrong direction wastes valuable tokens. A major bottleneck in agent performance is a lack of the right context, especially tacit knowledge that lives only inside human heads and cannot be found in written documents.

The Role of Feedback and UX

When agents tackle subjective tasks alongside humans, they need constant feedback to learn brand tastes, strategic preferences, and nuanced behaviors. Building an intuitive and delightful UX for reviewing agent work is essential. Since the bottleneck for agent performance is learning from feedback, optimizing the ease and quality of feedback collection unlocks long-horizon capabilities.

Case Study: Actively AI

Actively AI initially designed sales agents on a per-account basis, giving each agent the goal of moving a single prospective customer forward. However, they found that salespeople suffered from constant context switching and struggled to share learnings across different accounts. In their second iteration, Actively shifted to "motions", where agents handle specific jobs to be done, such as outbounding or deal progression, across multiple accounts. This aligned the workflow with how humans naturally review and provide feedback, allowing agents to learn much more quickly while background agents gather objective account data.