As we've built Aampe, I've come to the view that if you're trying to build or evaluate an agentic system, here are the questions you need to answer:
Do you have a massive, dynamic inventory of messages?
An agentic system requires thousands — often tens or hundreds of thousands — of unique, sendable messages. This isn't about blasting users, but about being equipped with the right content, framing, tone, etc. at the right moment, repeatedly. If bland or redundant wording obscures the value of what you're presenting to users, then agents will just learn noise, which will keep them in constant exploration.
Does your system reason over abstract concepts, not just individual messages?
You need an abstraction layer that captures commonalities across your inventory — value propositions, categories, incentives. These higher-order concepts are the real units of exploration and evaluation. If you only learn on performance of whole messages or message structure (e.g. subject lines), there's no analogical thinking, and therefore no good basis for choosing next steps. A system that can't choose its own next steps is not agentic.
Do you have a reward function that allows the agent to use every behavioral signal as feedback?
Messages should be evaluated not just on immediate conversions (and certainly not just on clicks), but on whether they advance the user toward a meaningful business goal. If your system can't identify that a user responded to a message in a way that made them more likely to convert in the future - even if they didn't actually convert after they got the message — then it's not going to have enough signal to calibrate it's choices.
Does the system track individual performance first and foremost?
To be agentic, a system must learn how different message attributes perform for each user, not just the average user. For any given user, an agent should be able to easily show how all Monday morning message performed, or how all all shoe messages emphasizing personal style performed. Yes, you should also be able to aggregate that up to see general trends, but the agent's view should be long, not wide.
Are you modeling performance as a distribution, not a point estimate?
Confidence matters. A single good outcome is enough basis for change agent behavior, but it shouldn't change agent behavior as much as 10 good outcomes. And even a single outcome should be given different priority depending on what the outcome looked like (see my point about reward functions, above). Agentic systems must understand and reason about impact as a probability distribution — weighing potential gains against the risks of being wrong.
If the system doesn't do these five things, I have a hard time seeing how it can operate agentically.
