
Voyager
@zrvmsummer
727 Following
13 Followers
0 reply
0 recast
0 reaction
0 reply
0 recast
0 reaction
0 reply
0 recast
0 reaction
0 reply
0 recast
0 reaction
0 reply
0 recast
0 reaction
0 reply
0 recast
0 reaction
5 replies
2 recasts
20 reactions
8 replies
1 recast
46 reactions
0 reply
0 recast
0 reaction
15 replies
3 recasts
63 reactions
0 reply
0 recast
0 reaction
17 replies
2 recasts
59 reactions
0 reply
0 recast
0 reaction
0 reply
0 recast
0 reaction
21 replies
4 recasts
39 reactions
on agent dev: sometimes a feature or bug fix is just adding another clause to the prompt, or fixing grammar.
It’s cool on one hand, that the prompt is a living document that’s both specification and implementation, but also clunky because English lacks the precision that a programming language has.
Because of this it’s also easy to introduce regressions because you don’t know how an llm will interpret changes to a prompt. Adding “IMPORTANT” might deemphasize some other rule, being too specific might make it dumb or less creative in other ways.
In code it’s deterministic, with llms it’s probabilistic.
So testing, aka evals, has become obviously very important, both for productivity and quality and doubly so if you’re handling natural language as input.
The actual agent code itself is quite trivial, prompts and functions, but having it work consistently and optimally for your input set is the bulk of the work, I think. 10 replies
3 recasts
36 reactions
11 replies
2 recasts
38 reactions
9 replies
0 recast
29 reactions
3 replies
3 recasts
30 reactions
9 replies
1 recast
47 reactions