Suggestion: Domain and dependency metadata for skills
In my personal experience building and working with agents, one feature I have found particularly valuable for skill discovery is adding, at minimum, explicit domain and dependencies metadata to skills.
As the skill set grows, providing the agent with the complete skill index can become increasingly expensive in terms of context budget. This becomes especially noticeable when only a small subset of skills is relevant to the task at hand.
On-demand skill discovery
A pattern that has worked very well for me is providing the agent with a very small core profile and a search_knowledge (or equivalent) tool, with an explicit instruction along the lines of:
When you lack the skills or knowledge required to perform a task, search the knowledge base for the relevant skills before proceeding.
If skills are indexed in a retrievable/vectorized knowledge base, this enables on-demand skill discovery rather than requiring the entire skill catalogue to be present in the context from the beginning.
For example, a skill could expose minimal metadata such as:
domain: astronomy
dependencies:
- image-processing
- coordinate-systems
The agent can then retrieve the relevant skill subset based on the task and progressively load what is actually required.
Dependencies as a discovery mechanism
This also leads naturally to the concept of specialized skill subsets.
A high-level skill may depend on a small number of more specialized skills, without requiring those dependencies to be included in the initial context. Once the parent skill is discovered, its dependencies can be resolved on demand.
This creates a lightweight dependency graph:
task
↓
domain discovery
↓
relevant skill
↓
dependencies
↓
specialized skills
The main benefit is that the model receives the indispensable context required for the action, rather than the entire universe of available skills.
I believe this could make the specification scale much better for agents with large and heterogeneous skill sets, while also making skill discovery more robust and efficient.
Possible minimal addition
Even without standardizing a full dependency-resolution mechanism, simply reserving standardized metadata fields such as:
would provide enough structure for implementations to build efficient discovery and retrieval mechanisms on top of the specification.
Suggestion: Domain and dependency metadata for skills
In my personal experience building and working with agents, one feature I have found particularly valuable for skill discovery is adding, at minimum, explicit domain and dependencies metadata to skills.
As the skill set grows, providing the agent with the complete skill index can become increasingly expensive in terms of context budget. This becomes especially noticeable when only a small subset of skills is relevant to the task at hand.
On-demand skill discovery
A pattern that has worked very well for me is providing the agent with a very small core profile and a
search_knowledge(or equivalent) tool, with an explicit instruction along the lines of:If skills are indexed in a retrievable/vectorized knowledge base, this enables on-demand skill discovery rather than requiring the entire skill catalogue to be present in the context from the beginning.
For example, a skill could expose minimal metadata such as:
The agent can then retrieve the relevant skill subset based on the task and progressively load what is actually required.
Dependencies as a discovery mechanism
This also leads naturally to the concept of specialized skill subsets.
A high-level skill may depend on a small number of more specialized skills, without requiring those dependencies to be included in the initial context. Once the parent skill is discovered, its dependencies can be resolved on demand.
This creates a lightweight dependency graph:
The main benefit is that the model receives the indispensable context required for the action, rather than the entire universe of available skills.
I believe this could make the specification scale much better for agents with large and heterogeneous skill sets, while also making skill discovery more robust and efficient.
Possible minimal addition
Even without standardizing a full dependency-resolution mechanism, simply reserving standardized metadata fields such as:
domaindependencieswould provide enough structure for implementations to build efficient discovery and retrieval mechanisms on top of the specification.