Triggers Live Inside the Workflow Now
Eyal distinguishes external triggers, which you send, from internal triggers, which fire from within the user's own routine. In an AI-native product, the external trigger is often an agent initiating contact — a summary delivered, a task completed, a flag raised. The agent's output becomes the cue. That means the trigger is only as reliable as the workflow producing it: if the agent misfires or halluccinates, the trigger arrives broken and the loop collapses before it starts. Designing the workflow tightly is designing the trigger.
Variable Reward Means Trustworthy Unpredictability
The reward in Eyal's model is variable precisely because certainty is boring — the unpredictable hit is what pulls users back. Agents introduce variability by default, which is a gift and a hazard simultaneously. The gift is genuine surprise: an agent surfacing a connection a human missed. The hazard is unreliable output eroding the reward entirely. Evaluation harnesses turn variability from noise into signal — they do not flatten the surprise, they eliminate the garbage so that what remains feels earned rather than random.
Investment Shapes the Agent, Not Just the Profile
In Hooked, the investment phase is where users load future value into the product — data, preferences, reputation — making tomorrow's trigger more accurate than today's. With agents, investment works the same way but runs deeper. Every correction a user makes, every boundary they set, every piece of context they provide trains the system's behavior going forward. That feedback loop is structural, not incidental; building it deliberately is what separates an AI-native company from a product with a chatbot bolted on.
Ethics Means Owning What Keeps People Returning
Eyal closes Hooked with a direct challenge: know whether you are a facilitator helping users reach their own goals, or a manipulator exploiting weak triggers. That line matters more when the agent acts on behalf of the user with real permissions and real consequences. Human judgment cannot be delegated away — someone must own the question of what the agent is actually optimizing for, whether the habit being formed is genuinely useful, and where the off-ramp lives. Accountability is not a feature to add later.
