§1. One of the advantages of an 'agent-native' architecture is that it is flexible enough to permit many use-cases. If, for example, pond's ingestion layer was agent-native, it'd offer the user to upload files of multiple formats and structures and describe how they'd like pond to interpret them. Over time, pond can recognise frequently recurring file-types, and perhaps design for itself an ingestion skill specifically for this user. The price one pays for the flexibility of this user experience of course is in the form of sacrifices to reliability and and increased LLM-usage cost.
§2. Pond is however at a stage at which it could seriously benefit from learning from and adapting to real user workflows. At first glance, I see two ways to approach this: (1) give the pond agent user-level (and user-mediated) access to pond's capabilities, and allow the user to interact with the agent to accomplish what they would like. Over time, as the agent collects records of sessions of interaction, we'll come to understand what exactly users use pond for. (2) more traditionally, onboard users that you expect are interested in pond's value, allow them to use it, and interview them to understand whether the product serves their needs.
§3. (1) is expensive, and it's not clear that even after I've identified standard usage pathways I can roll back agent capabilities and simply implement the product affordances that user interaction records suggest are most necessary. There's the further risk that offering general LLM capabilities in pond works against its stated goals of steering clear of finality—in other words, we introduce the affordance of asking the agent to synthesise or understand for us. Finally, pond has fairly strictly defined usage pathways; it's possible that this is a failing on its part, but it's possible also that these are constraints that make it meaningful for a particular kind of activity—that of understanding. An agent with user-level access to these pathways doesn't add all that much: it does not sit atop a general and extensive field of capabilities.