Another question, what are your thoughts if we also implemented “skills” primitives? I understand that we can expand “tools”; however, it may bring confusion, and “skills” may behave differently in the future vs current “tools” implementation, which could cause issues.
The current way we’re doing this for our agents is by allowing them to write files to a sandboxed context. Bash Tool lets them control their own sandboxed file system so they can write memories down to files and read them later. This is also an effective way to implement skills, letting the agent read the skill from the file system as it needs it just like the coding agents do
Ok, it looks like we do not need to implement skills as the ball is already rolling. I am assuming that bash-tool will be renamed to skills? I noticed that the terminology we use is non-traditional.
Regarding memory primitives. What I mean is memory related to an agent itself and an ability to connect/attach it to external Data Bases such as PostgreSQL, Weaviat, Neo4j (you name it). When I was looking at the library, I could not find those primitives. To me, it looks like a useful feature. Thoughts or suggestions?
On a slightly different topic. What’s the best way to ship new features? About 9 days ago, I was looking into opening another PR with skills, but I procrastinated as I was not sure what the best way to approach was without polluting PR threads.
Bash tool will not be renamed, it’s just a general tool for giving the agent safe access to bash with which it can manipulate files
It just turns out this is all agents need to use skills, so no custom ai sdk feature required
If you want something different with skill or memory handling you can make a new tool for it, but the existing ai sdk already makes this easy to implement in userland