Post #2365610

2026-02-23 06:03 UTC

@mpowney@mastodon.au I had the same experience, but I wonder if it was because the models I used.

Replies (1)

@mpowney@mastodon.au 2026-02-23 06:14

@florenciocano@infosec.exchange I suspect it’s not designed with much resilience or user feedback for use cases where lesser models than the ones typically used from Anthropic / OpenAI are configured. When I tried Azure Foundry hosted OpenAI models it was better Instead of digging in to why the local models didn’t work (the channel / web chat didn’t help and there was no relevant info in logs) I decided instead to build my own agent. The idea is to build it with more “robustness” including typescript class inheritance for key components like models, channels, skills, and agents. Also build in a containerisation layer so skills can be off loaded and isolated “properly”. Plus use different model configs for different discrete agent tasks eg. select your ideal model for a) planning b) skills parameter selection c) assessment of task completion and d) response to user. The theory with multiple model config is you could use a lightweight instruct model to work out which skills are relevant to a task, similarly for assessment, etc. models could also be fine tuned for the skills related tasks…

Open ##2365611