When Agent Goals Are Set Wrong, the Token Bill Gives Feedback First

Agent goal design is becoming an engineering problem. Harrison Chase of LangChain, together with mgarimella and Anshul Gupta of Actively AI, co-authored “Goal Engineering: How to Design Goals”; the original is on X. The article discusses a very specific point: when setting goals for agents, first think clearly about how humans will give feedback.

A bit of background. Agents have moved from text generation to tool calling, decision-making, and task execution, and next they will take on open-ended goals. If goal scope is set wrong, token consumption is only the first bill to appear. The article mentions that in Hone’s agent settings, one item is defining optimization metrics for the agent.

The article’s core judgment is: when designing agent scope, make it as easy as possible for humans to give high-quality feedback while the agent works within that scope. The reason is not complicated. Agent errors often come from missing the right context. Context such as documents and APIs is easy to add. The hardest part is knowledge that has not been written down, which can only be handed over by human experts through feedback. This becomes more obvious once agents start doing subjective work. For example, when optimizing marketing metrics, how provocative the brand can be, aesthetic preferences, and expression boundaries are all hard to put into documents. If the agent cannot quickly learn these from feedback, it will soon become useless on long tasks. Therefore, the UX for reviewing agent work is worth investing in. The bottleneck is often learning from feedback, so the convenience and quality of feedback collection need to be optimized.

The case comes from Actively AI. The first version built agents by account, with the goal of advancing accounts and producing actions for sales to approve. This seemed reasonable; account context seemed most important. After running it, they found another type of context was equally critical: how to do work efficiently. Specifically: outbound, replies, deal strategy, aggressiveness, and how to navigate large organizations. This knowledge was scattered across multiple agents. Outbound experience learned in one account often needed to be transferred to other accounts. Building agents by account made sales switch context frequently, and sharing learnings across agents was awkward.

The second version is called motions, based on Deep Agents. A motion is a specific job to be done, such as outbound or early-stage deal progression. A motion contains a bunch of similar tasks, and sales can review them together, with feedback transferring from one task to another. Account context is still important, but it is often already written down and more objective, assembled by background context-gathering agents, and pulled in on demand as motions advance the funnel across accounts.

The future judgment is also direct: humans reviewing agent work takes a lot of time, but it is high-value, because this is how domain knowledge is taught to agents. The goal is to align the agent’s goals with the granularity humans can most easily review. At the end of the article, Harrison Chase puts it very absolutely: prompt engineering and context engineering will disappear; goal engineering will remain. This judgment can be debated, but it points to a reality: goal scope design directly affects feedback quality, learning speed, and whether an agent can remain autonomous over the long term.

Original article: https://x.com/i/article/2121382906820243456

Original tweet: https://x.com/Hacubu/status/2121665217809801615