#Can i make agent more effective with tooling hints
1 messages · Page 1 of 1 (latest)
another one today
44 : ┆ select_methods(methods: ["file", "directory", "withNewFile", "withExec", "withWorkdir"]): String!
44 : ┆ select_methods ERROR [0.0s]
44 : ┆ ! unknown methods: [file directory withNewFile withExec withWorkdir]; use list_methods first
@light cliff can you remind me, are you using LLM or dagger mcpor prompt mode in dagger shell?
Do you have a module with tools named similar to the Dagger API or is that the Dagger API?
i'm using dagger call in azure pipelines. No interactive usage yet. Haven't tried dagger shell, crush, or any fancy new agentic type cli tools yet cause I've been heads down on just getting this CI spotless 🙂
a little messy as I've been working through it but will probably add private descriptors and more to cleanup soon 🙂
Nice! I think to Jeremy's question it sounds like you're using dag.LLM based on the select_methods/list_methods output. By default that will include a system prompt that instructs the LLM on how to use list/select (https://github.com/dagger/dagger/blob/main/core/llm_dagger_prompt.md) and also the function + parameter descriptions are available to the LLM automatically. I will add that most models will still get a bit lost with dagger's core API and trying to chain objects + functions, so in my experience its been better to give the LLM custom objects that don't expose core types
An open-source runtime for composable workflows. Great for AI agents and CI/CD. - dagger/dagger
Might experiment with WithSystemPrompt then as I see I can override, and add that + some directions in to see if changes behavior.
nice, yeah WithSystemPrompt is additive, so you can add as much guidance as you want there. Or if you want to try from a blank slate you can do WithoutDefaultSystemPrompt first