Harrison Chase: Prompt Engineering Will Disappear, Goal Engineering Won't
LangChain's Harrison Chase and the Actively AI team discuss agent goal design: the scope of goals should make it easy for humans to continuously give feedback. Actively changed its sales agent from being built around accounts to being built around sales motions, so feedback can finally be reused across tasks.
Setting goals for an agent is most easily reduced to a single prompt. In the article, LangChain's Harrison Chase and the Actively AI team break this down more carefully: the scope of a goal should be designed around whether humans can conveniently give high-quality feedback.
The original text first provides some background. The scope of agent tasks has expanded rapidly in recent years, from text generation, tool calling, and decision-making to executing complete tasks. The next step is to take on open-ended goals and push toward completion on their own. Products like Hone show the direction. In Hone's agent settings, a key action is defining optimization metrics for the agent. Point the goal in the wrong direction, and tokens are burned for nothing.
Why has feedback become central? When agents make mistakes, a large part of the reason is missing context. Context such as documents and APIs is relatively easy to add. The hardest thing to add is what has not been written down and exists only in experts' heads. It can only come from humans giving feedback while the agent works. This becomes even more obvious once agents start taking on subjective work. For example, when optimizing ad campaigns, besides factual information, there is brand tone, how provocative it can be, and what counts as good-looking. If an agent cannot learn these things, it quickly becomes useless.
So when designing the scope of a goal, feedback should be easy to produce. The UX for reviewing agent work is worth investing in. The original text says the bottleneck in agent performance is often learning from feedback. Optimizing the efficiency and quality of feedback collection has a high return.
The case is Actively AI, which builds agents for sales teams. The first version built an agent per account, one agent per prospect, with the goal of advancing the account and producing actions for sales to approve. It sounds reasonable; account context is the most important. Once it was running, they found the agent also needed another kind of context: how to do this work efficiently. How to do outbound, how to reply, how to do deal strategy, how aggressive to be, how to navigate a large organization. This knowledge was scattered across multiple agents, and what was learned from one account could often transfer to other accounts.
The problem with building agents by account is feedback. When a salesperson is looking at a single account, there is only a small amount to review at a time, and they frequently switch between agents. Sharing experience across agents is also awkward. The second version changed to motions, built on Deep Agents. A motion is a type of work, such as outbound or early deal advancement. The same type of task is reviewed together; a salesperson gives feedback once, and it naturally transfers to similar tasks. Account context is still important, but most of it has already been written down and is more objective. Actively uses background context-gathering agents to aggregate account context, and motion agents pull it as needed.
Some commenters put this more directly: motions are just the division of labor that sales teams already have; outbound and deals each own a stage. Goal engineering goes around in a circle and ends up replicating the org chart. This observation makes sense. Human teams divide by work because that is where feedback happens. Agents have grown into the same shape.
The risks are also in the comments. Someone says that giving an agent a vague goal is the fastest way to burn tokens. Someone says most agent goals are like New Year's resolutions: ambitious, and abandoned by February. Others mention Goodhart's law: humans will bend a bad goal a little, while agents will execute it all the way through. Give only "bring in 100 new customers" without setting brand boundaries, and the agent may mass-message 10,000 people and burn your reputation. The more singular the goal is written, the more the agent will take the literal meaning and push the action to the extreme.
Harrison Chase's judgment is blunt: prompt engineering and context engineering will disappear; goal engineering will remain. Future agents will receive a goal and optimize it themselves. Human review takes time, but it is valuable. That is how domain knowledge is taught to agents. Agent goals should be aligned with what humans can most easily review, so feedback can be extracted, and only then can agents keep learning, becoming more useful and more autonomous.
Original link: https://x.com/hwchase17/status/2121635683999883577
Article link: https://x.com/i/article/2121382906820243456
发布时间: 2026-11-15 15:30