Skip to main content

​
What Is a “Good” Event Feed?

Agentic learning systems depend on high-quality behavioral data to evaluate the impact of their actions, yet teams often misunderstand which data are most useful and why. This paper outlines the categories of data relevant to agentic systems, explains the structural requirements for meaningful learning, and offers a practical framework for prioritizing event instrumentation based on their contribution to agent learning and decision-making.

​
Categories of Data in Agentic Systems

Agentic systems are designed to learn by observing the effects of their actions on users. To facilitate that learning, they rely on three fundamental categories of data: user attributes, content attributes, and timestamped events. Each of these serves a distinct role in the agent’s learning environment.
  • User attributes include information such as location, language, subscription tier, or behavioral rollups like “total streams in the last 7 days.” These attributes are inherently static and often slow to change, and they lack timestamp precision. This means they cannot be used to observe the direct impact of a treatment. They can be used to filter user eligibility or match similar users for learning acceleration, which is important functionality, but it’s functionality that facilitates rather than drives learning.
  • Content attributes like genre, artist, brand, or feature category are descriptive metadata about items that users might interact with or that are referenced in messages. They are vital to content personalization, allowing systems to craft highly relevant treatments, sometimes via recommender systems. Still, these attributes pertain to the treatments themselves rather than their outcomes. They tell us what was recommended or referenced, but not how the user responded.
  • Timestamped user events are the most valuable for agentic learning. These are the actions a user takes at a specific moment in time - starting a stream, completing a purchase, reading an article, logging a workout. Because they are temporally situated, timestamped events allow the system to draw causal inferences.

​
Agentic Learning Needs Timestamped Events More Than Any Other Kind of Data

The nature of agentic learning requires the ability to isolate the effect of individual treatments. Agents must track when treatments are delivered, to whom, what the content was, and what happened afterward. A timestamped event is the only way to make that linkage robust. If those events are absent, or if they exist only as coarse or delayed aggregates, the agent loses its capacity to learn precise and timely cause-effect relationships.  Not all events are equally valuable. Some user actions serve as excellent signals for learning because they represent volitional, meaningful, and spontaneous engagement. Others, while perhaps business-critical, fail to serve as reliable learning targets. Subscription renewals, for example, are predictable and scheduled. They are certainly important but do not serve as good reward events because they don’t indicate that a user took action in response to a treatment. Similarly, actions that are finite or cannot be repeated—such as linking an account—do not provide the ongoing feedback loop that agentic systems rely on. Quantity matters alongside quality. Even if the most valuable events are identified, restricting agents to too narrow a set leaves them starved of signal. A richer variety of timestamped events increases the chances of capturing subtle and distributed forms of user response, allowing agents to triangulate impact more reliably. When the event stream is too sparse, agents are prone to treating missing data as non-response, creating the illusion that their actions had no effect. Broad event coverage ensures that genuine patterns of influence are visible rather than hidden in unobserved behavior.

​
Characteristics of Good Reward Events

Good reward events are distinguished by a set of properties that make them reliable indicators of user intent and value alignment. These events provide the backbone for agentic learning because they allow the system to observe how users react to treatments in ways that are both timely and meaningful. Not every user action qualifies. Some are noisy, passive, or structurally unsuitable for learning. The most effective reward events share these qualities:
  • Volitional: The user initiates the action by choice, not passively or automatically.
  • Spontaneous: The event can occur at any time; it is not bound to a predefined schedule.
  • Repeatable: The user can perform the action multiple times without exhausting the opportunity.
  • Value-Aligned: The event reflects real usage of the product or service in a way that supports the business’s long-term goals.
These traits ensure that the system learns from actions that are not only detectable but also meaningful. For instance, a user streaming a song, purchasing a product, or completing a fitness goal are all behaviors that signal engagement and satisfaction. They are voluntary, time-sensitive, and repeatable - and they directly express how well the product is delivering value. By contrast, events like subscription renewals or account linkages may be strategically important, but they occur on predictable schedules or as one-time actions, making them poor choices for continuous feedback and learning. Systems that treat such events as reward signals risk mistaking coincidence for causality and may end up optimizing toward the wrong kinds of behaviors.

​
A Checklist for Event Prioritization

To assist product and data teams in determining which events to prioritize, the following checklist ranks event types by their usefulness to agentic learning. Each tier represents a broad category of events, with a description of how those events contribute to learning. Teams should aim to provide as many as possible, starting from the top.
  • Tier 1: Core Value Actions. These events represent the most direct indicators of meaningful, volitional use of a product’s core offering. Examples include content streamed or consumed (such as music, video, or articles), products purchased, transactions completed (such as bill payments or fund transfers), and other key actions logged.
  • Tier 2: Feature Engagement. Events in this tier reflect deliberate interaction with features that support or extend core value. These include actions like adding items to playlists, wishlists, or carts, toggling specific settings or features, and modifying subscription plans. They offer strong signals of intentional product use.
  • Tier 3: Session and Message Engagement. These events indicate that the user was present and possibly attentive, including app opens, session starts and stops, notification clicks, and message views. They help establish temporal windows for attribution and support models of general engagement, though they are more ambiguous as indicators of product value.
  • Tier 4: Utility Interactions. This tier includes search events and navigational actions that show intent but are not strongly linked to product consumption. General site search, browsing, and tag-based exploration fall into this category. These actions are often preliminary and may precede more meaningful behaviors but are less discriminative in isolation.
  • Tier 5: Passive Interaction Events. These are events like generic page views or button clicks that are not tied to specific features or content. They are useful for understanding interface flow and may support debugging or UI experiments, but they carry limited value for modeling user satisfaction or learning preference structures.
  • Tier 6: Scheduled or One-Time Outcomes. Events such as subscription renewals, trial completions, or account linkages often reflect important business milestones, but they are either externally timed or terminal. As such, they are poor indicators for causal modeling and are unsuitable as reward signals.
Note that events that make very poor reward signal may still be very important to the business. Agents generally should not try to learn user preferences based on schedule subscription renewals, but those renewals can and should be tracked over time, since better, more meaningful, and more frequent engagement with the core value of the platform should correlate with increased retention and renewal over time. Agents can support these business priorities without needing to learn directly on the events associated with those priorities.

​
Avoiding Unnecessary Data Volume

For many organizations - particularly those operating in event-driven architectures such as streaming services - the volume of raw behavioral data can be immense. Ingesting and transmitting all raw playback events, particularly if sampled at regular time intervals, is both financially and technically infeasible. Fortunately, agentic systems do not require full-resolution event streams to learn effectively. The key is to identify compact and representative signals. For streaming apps, a single “stream start” event per session or item is typically sufficient. This event can be defined as the first timestamp at which a user begins interacting with a new item - meaning an item ID that differs from the one involved in the most recent stream event. This approach preserves the essential behavioral insight without overwhelming infrastructure. A corresponding “stream stop” event could also be useful and meaningful. The broader principle is that agentic learning depends on timing, not exhaustive detail. A sparse set of carefully selected, timestamped events offers more actionable learning signal than high-volume logs that are costly to transmit and difficult to operationalize. For domains where data transfer costs are high or real-time telemetry is impractical, synthesized summary events - so long as they retain specificity and timestamp precision - can strike the right balance between learning fidelity and system efficiency.

​
Timestamped Behavior Is the Foundation of Agentic Learning

Effective agentic learning depends on timestamped behavioral data. These are the only data that allow agents to evaluate the success or failure of their actions. User and content attributes have their place, particularly in shaping treatments and filtering eligibility, but they do not enable learning by themselves. Reward events must be immediate, repeatable, and volitional. Systems that lack this data are limited in what they can optimize for and risk training agents on misleading or irrelevant signals. For teams deciding what event data to expose to an agentic system, the clearest priority is to surface the events that reflect ongoing engagement with the app’s core value proposition - and to do so with as much temporal fidelity as possible.
⌘I