Hierarchical Skills

[ron-clover]

Hi MAF Team ! :)

Problem Statement

I've been using Agent Skills and I think they are great.
Progressive Disclosure is an awesome way to manage the harness.

However, once a repo contains 100+ skills (which could happen in enterprise solutions) - it gets hard to manage.
Issues like:

• Boundaries between skills start to get blurry
• Context is polluted with hundreds of descriptions before the chat even started.

As far as I understand, ALL skills frontmatters are currently loaded automatically to the context window.

Ths solution

I suggest supporting a hierarchical structure of "skill trees".
You would start by disclosing "root skills" - these are skills that are introduced in every turn.

Then, as the agent loads a skill - this skill might disclose related sub-skills frontmatters.
Thus the agent can decide to continue and load this sub skill if it wants.

Note that each of those sub skills can have their own scripts and resources - providing a robust hirarchical strcuture.

Example

Say we have a guitar skill - it explains what a guitar is.
Then, this skill will have the following sub-skills - chords skill, and scales skill.

It makes a lot of sense that only once the agent decided to learn guitar - only then will the more granular steps be introduced.

Lmk what you think - I think this would be great for the framework.

Thx,
Ron

[ron-clover]

Another option is to just allow the users to control which skills get introduced by default - and which will just get introduced int he prompt of other skills..

I guess this is much simpler to implement

[abhishekabramaina]

@ron-clover I understood what you are talking, Progressive Disclosure!. and Yes, It makes much more sense to load skills progressively rather than bloating up the context with descriptions of every skill that exist.

[moonbox3]

@SergeyMenshykh any thoughts here?