So, I've been noticing a few key problems with agent integration in current setups.
1. Most software backends are built in non-Pythonic languages (Rust, Golang, Java, etc.). But agents are mostly in Python, with some having JS support max. How do we integrate them efficiently? Basically, cross-language integration; right? But how to do that without major headaches?
2. If the agent developer creates an API wrapper for it, the same process has to be followed for integrating another agent. Then third, fourth etc etc.
3. Reliable agent integration in current software often requires not just open monitoring but also tracking how the agent responds in day-to-day scenarios. For this, we need memory, right? But not the typical memory that remembers specific moments. Instead, the agent should learn how to interact with the customer. This behavioral improvement we're calling Action Memory, where instead of focusing on what to remember, it emphasizes how to remember.
4. If we accept that agents are the next big thing after LLMs, then we have to focus on efficient execution of those agents too. For this, we need to lean into the serverless side but ensure compatibility with persistent agents.Focusing on all these things, we're building RunAgent.