Skip to content
Goal Engineering: Aligning AI Agent Goals with Human Feedback Loops

Goal Engineering: Aligning AI Agent Goals with Human Feedback Loops

The article introduces goal engineering for AI agents, arguing that effective goal design must align with how humans naturally provide feedback, using Actively AI's transition from per-customer agents to reusable 'motions' as a case study to show how this enables scalable learning and continuous improvement.

ThinkInAI Community
ThinkInAI Community
ThinkInAI Community

Over the past few years, the range of tasks AI agents can handle has expanded dramatically. They have evolved from simple text generation to calling tools, making decisions, and executing complete tasks. The future of agents lies in taking on open-ended goals and persistently driving toward completion. Products like Hone (https://x.com/hone/status/2108218574926176490?s=20) offer a glimpse of that future.

In such a world, defining the scope of an agent's work is critical. Pointing agents in the wrong direction wastes massive amounts of tokens. A key step in configuring Hone agents is defining the metrics they should optimize.

How to Design Goals

When studying similar agents, we realized that designing an agent's work scope should make it as easy as possible for humans to give feedback on the agent's work within that scope.

Why? A major reason agents fail is missing the right context. Solving this requires providing appropriate context. Some context — like access to the right documents or APIs — is relatively easy to supply. The hardest context to provide is the kind that was never documented in the first place. That undocumented context must be passed from human experts to the agent through feedback.

This becomes especially important as agents move beyond verifiable tasks and start doing subjective, organizational work alongside humans. For example, if you want an agent to improve a business metric such as optimizing ads for better marketing performance, there are many subtle constraints beyond factual context. You may have specific judgments about what copy or visuals count as good, and clear opinions on how bold or edgy the brand can be. If the agent cannot efficiently absorb that feedback and learn those nuances, it quickly becomes useless.

Article illustration showing feedback loop concepts
Article illustration showing feedback loop concepts

If an agent's work scope is clear enough that humans can easily provide large amounts of high-quality feedback, that dramatically helps build an agent that can reliably pursue long-term goals.

A related lesson: you need to invest heavily in building an intuitive, delightful user experience for reviewing agent work. The bottleneck for agent improvement is often whether they can learn from feedback. Therefore, you should optimize the ease and quality of feedback collection. Investing resources to craft a clear review experience is well worth it.

Case Study: Actively AI

Actively AI (https://www.actively.ai/) builds agents for sales teams.

Actively's first goal-oriented agent experience created one agent per prospect — where a prospect is a potential customer the sales team wants to close. The agent's job was to advance the prospect and generate actions for the sales rep to approve. Intuitively, this design made sense: it was a relatively constrained, well-defined goal with a clear outcome. At first glance, prospect-specific context seemed like the most important context, so organizing agents around prospects appeared logical.

As the agents started working, the team discovered another critical class of context: how to do the job effectively. This includes how to conduct outreach, how to reply, how to strategize deals, how aggressive to be, and how to navigate large organizations. This context is scattered across multiple agents: if you learn an outreach lesson on one prospect, you usually want to transfer that lesson to other prospects.

Actively found that when agents focused on a single prospect, salespeople struggled to use them efficiently and struggled to give feedback. An agent advancing a single prospect must perform many different tasks, and those tasks are often spread out over time. Two problems emerged: (a) salespeople had to context-switch across many agents because any single agent only had a few items ready for review at a given moment; (b) sharing learnings across agents was awkward, yet necessary because those agents were learning the same kinds of things.

Diagram comparing per-prospect vs motion-based agent organization
Diagram comparing per-prospect vs motion-based agent organization

Actively's second version solved these problems. Their new goal-oriented agents are built on Deep Agents (https://github.com/langchain-ai/deepagents) and are called "motions." A motion represents a specific unit of work, such as conducting outreach or advancing early-stage deals. How to complete that work often has no single right answer; much of it depends on each sales rep's personal preferences. Salespeople hold vast tacit knowledge that must be fed back to the agent owning that motion so it continuously learns. Because a motion groups similar tasks, salespeople can easily review those tasks together. Any feedback from one task intuitively transfers to other tasks within the same motion.

This design keeps feedback aligned with the agent — both in scope of applicability and in ease of delivery — making it easier to build powerful agents.

This is not to say prospect context is unimportant; it certainly is. But prospect context is often already documented and more objective. So Actively lets context-gathering agents run in the background, compiling each prospect's information, while the actual work of moving prospects through the funnel is handled by motions that operate across many prospects. Motions own a specific stage of the sales process, learn faster from feedback, and pull in prospect context when needed.

The Future

In the future, agents will reach a point where you simply give them a goal and they act autonomously, continuously improving how that goal is achieved.

Human review of agent work takes significant time, but that work is immensely valuable. It is how we teach agents to operate and adapt to a specific domain. The knowledge that lives only in people's heads is critical; it must be extracted and converted into a form that can be efficiently passed to agents.

To do this best, you need to align the agent's goals with what humans find easiest to review. That creates a smooth review loop, maximizes information extraction, and lets agents keep learning — becoming more useful and more autonomous over time.

Prompt engineering and context engineering will eventually fade, but goal engineering will remain.

Original Source

Signed-in readers can open the original source through BestHub's protected redirect.

Sign in to view source
Republication Notice

This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactadmin@besthub.devand we will review it promptly.

AI agentsAgent DesignSales Automationcontext engineeringDeep AgentsGoal EngineeringActively AIHuman Feedback Loops
Reader feedback

How this landed with the community

Sign in to like

Rate this article

Was this worth your time?

Sign in to rate
Discussion

0 Comments

Thoughtful readers leave field notes, pushback, and hard-won operational detail here.