artifacts that plug into a system

anthropic recently published agent skills as open standard.
a good thing about skills is that they are external to the system(llm/ai). this means they can not only be integrated with the system, but that they can also improved independently of the system.

a skill for performing a task 'y' can be iterated upon, with its historical performance as a feedback.

the same principle should be applied when frameworks for agent, and workflow orchestration. the definition should have lose coupling with the library handling the execution.

for instance, imagine a library called flow
typically such a library would expect it user to construct a workflow or agent as :

flow = new flow()
flow.add_agent()
flow.add_tool()
flow.add_prompt()
flow.build_system()

flow.run()

alternatively, if the library treated the workflow as independent artifact, it could simply do the following

flow = new flow()
flow.run()

while option a is more programmatic and code appropriate, its easier and somewhat cleaner to iterate on the workflow in case of option b.

and then the code evolves in a cleaner fashion to :

flow = new flow()
flow.run(flow_v2)

or we could just maintain the concept of 'latest' with the ability to rollback whenever required.

references

1. agent skill