I was curious about what are the variations possible for implementing the progressively expanding toolkit.
One thought that comes to mind is to have a sub agent continuously evaluating the conversation and deciding if new tools need to be activated but I guess a secondary challenge with this would be to then inject the tool definition into the context and that too in a way where it is retained for longer multi turn conversations. I guess this would make use of the tiered memory modeling that was discussed in the post.
The explore-plan-act loop pattern is underrated. I added an explicit planning phase to my agent about two months in and the quality of edits went up noticeably - just from separating the reading and writing phases. The lifecycle hooks section also resonates. I run hooks for session start, session end, and after any file modification.
Not because the model forgets procedural steps exactly, but because it's inconsistent under load. Hooks remove that inconsistency. The tiered memory pattern is something I implemented a bit differently - I call them working memory vs persistent memory - but same idea. Stuff that changes daily vs stuff that never changes belongs in different files.
good article 👍
This was a good read
This was a really good read!
I was curious about what are the variations possible for implementing the progressively expanding toolkit.
One thought that comes to mind is to have a sub agent continuously evaluating the conversation and deciding if new tools need to be activated but I guess a secondary challenge with this would be to then inject the tool definition into the context and that too in a way where it is retained for longer multi turn conversations. I guess this would make use of the tiered memory modeling that was discussed in the post.
The explore-plan-act loop pattern is underrated. I added an explicit planning phase to my agent about two months in and the quality of edits went up noticeably - just from separating the reading and writing phases. The lifecycle hooks section also resonates. I run hooks for session start, session end, and after any file modification.
Not because the model forgets procedural steps exactly, but because it's inconsistent under load. Hooks remove that inconsistency. The tiered memory pattern is something I implemented a bit differently - I call them working memory vs persistent memory - but same idea. Stuff that changes daily vs stuff that never changes belongs in different files.