The idea of treating AI agents as a distinct persona, similar to users or developers, is spot on. We really have to think about how they experience our software, this is essential for building effective systems that actually work for both humans and agents.
That being said I think the emphasis on APIs might be wrong in the long term. I think having an intake agent sitting on a side channel apart from HTTP/HTML on your product waiting to interact with external agents might be a better analogy. Maybe using just a mix of free text and fixed schemas with access to the product tools, who knows. Really excited to see developments on this area.
Yeah interestingly API's in their current form are rarely very good for agents. In many cases tools like Operator using a virtual browser and screenshotting are better for agent interactions than API specs.
This shows we need to build better approaches to agent interactions that are not at the level of "run a virtual browser", but that encodes much more of the workflows available than raw API's do today.
Related: Zdenek Nemec's talk "APIs for AI: Have we failed?" from Nordic APIs 2024 is an insightful articulation of AI agents as a distinct persona to accommodate in platform and security decisions.
That being said I think the emphasis on APIs might be wrong in the long term. I think having an intake agent sitting on a side channel apart from HTTP/HTML on your product waiting to interact with external agents might be a better analogy. Maybe using just a mix of free text and fixed schemas with access to the product tools, who knows. Really excited to see developments on this area.