I like what you did with the action catalog as a way to express bounded outcomes!
I think the biggest similar concept in our approaches is that operational safety really requires strict determinism in a lot of places. It's not like a coding agent where you can have it write itself shell scripts to get the job done any which way, and it'll probably be fine. So you wind up wrapping the probabilistic agent as a glue point between deterministic steps to get safe, reliable results.
It's great to be able to put automation into those places that would have required a human brain's worth of context before, because you'd never have been able to write a big enough if-case. But you need that deterministic input and outcome on each side to make sure that results are predictable enough that the system doesn't eat it.
<Insert lenghty ranting about "Go read In Search of Certainty again, Mark Burgess is right" here>
This is really a nice tool. For parking oversized tool results on disk, how are you storing them? Is that in flat files by byte-range reads, or embedding a SQLite or RocksDB to manage cached tool outputs?
I'm also one of the people working on this project.
Oversized result artifacts are session-namespaced files cached on disc. Coordinator runs are given a lookup manifest and the internal scratchpad tool can search, grep, etc. to find only the pieces they need with inner loops. This enables a sort of model-driven targeted context engineering.
Nice work. The part that collapses in IR agents is the permitted tool use enforced outside the model context.
Great looking project! I've been fumbling towards a similar approach in GoLang, http://github.com/ianeff/thump. I like yours better!
I like what you did with the action catalog as a way to express bounded outcomes!
I think the biggest similar concept in our approaches is that operational safety really requires strict determinism in a lot of places. It's not like a coding agent where you can have it write itself shell scripts to get the job done any which way, and it'll probably be fine. So you wind up wrapping the probabilistic agent as a glue point between deterministic steps to get safe, reliable results.
It's great to be able to put automation into those places that would have required a human brain's worth of context before, because you'd never have been able to write a big enough if-case. But you need that deterministic input and outcome on each side to make sure that results are predictable enough that the system doesn't eat it.
<Insert lenghty ranting about "Go read In Search of Certainty again, Mark Burgess is right" here>
This is really a nice tool. For parking oversized tool results on disk, how are you storing them? Is that in flat files by byte-range reads, or embedding a SQLite or RocksDB to manage cached tool outputs?
I'm also one of the people working on this project.
Oversized result artifacts are session-namespaced files cached on disc. Coordinator runs are given a lookup manifest and the internal scratchpad tool can search, grep, etc. to find only the pieces they need with inner loops. This enables a sort of model-driven targeted context engineering.
You can explore the implementation here: https://github.com/mezmo/aura/tree/main/crates/aura/src/scra...