Prompt engineering will disappear; the first hurdle of goal engineering is making it easy for humans to spot errors

LangChain's Harrison Chase and Actively AI's mgarimella wrote an article about goal engineering. Original link here: Goal Engineering: How to Design Goals. The core judgment is straightforward: when designing agent goals, first check whether humans can easily give high-quality feedback. Feedback is the hardest thing to extract for aligning agents with human mental models, and it is also what goal design most easily overlooks.

A common reason agents fail is missing context. Context such as docs and APIs is easy to add. What's hard to add is the judgment in experts' heads: what tone ad copy should have, how bold the brand can be, how aggressive an outbound email should be. None of this is written down anywhere; it can only come from humans reviewing and editing, then feeding it back to the agent. So goal scope and review UX need to be designed together. The article says that if an agent's scope allows humans to give high-quality feedback frequently, that agent is more likely to reliably pursue long-horizon goals.

The Actively AI case is concrete. The first version sliced by account: one agent was responsible for one prospect and produced actions for sales to approve. It sounded reasonable. In practice, sales could review only a few things at once and had to switch back and forth among multiple agents. Worse, outbound lessons learned in one account were hard to transfer to other accounts. The second version changed to slicing by motion. A motion is a category of work, such as outbound or early deal progression. One motion spans multiple accounts, so similar tasks can be reviewed in batches, and feedback is naturally reusable. Account context still matters; it is just delegated to a background agent to collect.

Image source: Harrison Chase tweet.

People in the replies added a few sharp points. John Rood said what really constrains an agent in a goal is what it cannot do; if you give only a goal, the agent will optimize literally and end up completing it technically. Kawai said vague goals burn tokens fastest. Amir said most agent goals are like New Year's resolutions: very ambitious, then abandoned in February. Victor said motions resemble the division of labor that sales teams already have, and a lot of goal engineering ends up copying the org chart. These comments are blunt, but they all point to the same thing: goals must be reviewable, editable, and reusable by humans.

The original article even makes a very strong claim: prompt engineering and context engineering will disappear, and goal engineering will remain. That claim can still be debated, but it points to a reality. Once agents take on open-ended goals, the human feedback interface is part of the system. Goals should become the entry point for humans to spot errors, not pretty words written for agents to read.

Related links:
Harrison Chase's tweet
Original source tweet