anthropic recently launched claude skills, which has the potentially to substantially simplify the implementation of this toolbox.
a catch is that skills are only supported for claude models. the simplest solution would be to drop support for all non-anthropic platforms, and open a new issue to add support back in at a future date, when other platforms (hopefully) adopt support for skills
this greatly simplifies many aspects of the toolbox:
- model registry: contains only anthropic models: opus 4 handles the most complex tasks; sonnet 4.5 provides a sensible default, and haiku 4.5 can do simple tasks cheaply.
- tools registry: now just a wrapper and interface for anthropic skills. the example skills repo contains a good base set of skills that we can directly copy. importantly, one of the "meta" skills included is a tool for creating new skills! so we can call on this if an appropriate skill isn't available.
- credentials management: only need support for anthropic API keys
- cost estimation: only need support for anthropic models
- local resource estimation: no longer needed, since all models run on anthropic's servers
- parallel tasks can be launched as parallel agents
so we can do a total refactor to cut away nearly all of the toolbox code, and substantially simplify the remaining code. the key will be do develop a reliable core set of functions to compile the pipeline (this must be done and managed automatically-- pipelines should be compiled to langgraphs) and then called.
anthropic recently launched claude skills, which has the potentially to substantially simplify the implementation of this toolbox.
a catch is that skills are only supported for claude models. the simplest solution would be to drop support for all non-anthropic platforms, and open a new issue to add support back in at a future date, when other platforms (hopefully) adopt support for skills
this greatly simplifies many aspects of the toolbox:
so we can do a total refactor to cut away nearly all of the toolbox code, and substantially simplify the remaining code. the key will be do develop a reliable core set of functions to compile the pipeline (this must be done and managed automatically-- pipelines should be compiled to langgraphs) and then called.