verbiricha on Nostr: oh, this is vastly different from my approach. i've opted for defining constrained ...

oh, this is vastly different from my approach.

i've opted for defining constrained workflows as a DAG with minimal control flow features: bounded loops and parallel fan out. i have one that does work (https://github.com/purrgrammer/fragua/blob/main/.fragua/workflows/work.yaml) and another that does review (https://github.com/purrgrammer/fragua/blob/main/.fragua/workflows/review.yaml) and i run these locally until they converge, then PR. CI also runs the review workflow, and periodic ones such as doc drift detection.

i'm able to schedule these to run periodically, so they can query the repo state or a DB and trigger other workflows.

the reason i'm doing it this way is predictability and observability. at $work i've implemented a few of them and i can use our existing PR corpus to evaluate. i'm applying this to the SDLC where the processes can easily translate to well-defined workflows, no need for multi-agent systems imo. n agents working on something is nondeterminismˆn.

fine-grained control of model allows to use the right model for the task (lowest tier models are decent classifiers and cheap af, want to use a strong model for planning and a cheaper one for executing, etc) and ability to run deterministic steps (scripts, commands) is also invaluable, i only reach for llms when there is genuine need for judgement.