#My personal coding agent 🧵

1 messages · Page 1 of 1 (latest)

merry wadi
#

@loud spade @swift crow @karmic egret @solid oyster OK so this is what I want to demo tonight 🙂 And more importantly it's the most promising use case for Dagger in the short term.

The key is to have a collection of modules that are actually useful, and the LLM can use to actually help me.

#

So...

  1. which modules are in our collection so far? 🙂
#
  • before you say anything @loud spade : weather 😛
  • github integration (for issues, pull requests)
  • ability to run the tests- but how to run the tests on a local dev version?
  • dev env
#
  1. assuming the standard workflow is: dagger init in your project. Then dagger install modules you want to use.

--> What's the example project?

loud spade
#

The hello-dagger project might be a bit boring but maybe it's simple enough to add some wacky things easily?

karmic egret
#

(I've been having mixed results in my API Agent using it)

merry wadi
#

What's the URl for hello-dagger ?

#

Can we collaborate on a repo that becomes our showcase project?

#

then I'll just demo from that repo tonight

karmic egret
merry wadi
#

ah yes

#

what does the app do?

#

webapp that says hello I guess

loud spade
#

Yup exactly. It should be a small enough context that any LLM should be able to figure it out and write code without too much hand holding

#

But if it's too boring we can come up with a different starting point

merry wadi
#

in this set up, the demo is only as cool as the modules we install

#

hello-dagger is good

#

wish we had that stdlib right now

merry wadi
#

@karmic egret @loud spade is there a branch where hello-dagger is daggerized?

loud spade
#

Like a single copy/paste

merry wadi
#

OK I demoed this flow yesterday. It was janky and over-complicated, but I think the fundamentals are strong. Would have been better with modules... but didn't have time to get to that

#

@swift crow request... if we want a CLI-first experience, it means your first agent is the CLI's builtin agent (increasingly I think we should just call it "Dagger Agent"). But then, how do you bridge from that builtin agent to custom agents written in code. with dag.LLM?

swift crow
#

I think that just comes down to overriding the system prompt and being able to control its tool calling scheme (either via single-object or MCP or [...])

merry wadi
#

So far you have 2 options:

  1. Eject out of CLI agent completely. Write your own, call it from the shell. No "prompt mode" involved. Jarring / disjointed.
  2. Multi-agent. CLI agent calls your custom agents as tools, via module composition. Cool but very advanced
#

My request is for option 3: customize the CLI agent(s) by setting variables of type LLM. Then allow cycling through them with the CLI keybind

#

Even cooler if the variable name becomes the agent name in the prompt 😛

swift crow
#

yea cool

#

we'll just have to be sure not to sync those variables to any LLM 😛

#

since if it swaps to them it'll break (>128 tools)

merry wadi
# 
thinker=$(llm --model=deepseek | with-prompt "you are a deep thinker. if I ask you for things, reply with a deep sigh, then think about it real hard. Then answer 42.")
junior=$(llm --model=gpt-3.5-turbo)