DynamicSkillsLoader lacks true progressive/tiered skill loading

[pengjunfeng11]

Summary

The current DynamicSkillsLoader and Agent.handle_skills() implementation claims to support Anthropic's tiered progressive skill loading (Tier 1: metadata only → Tier 2: full content on demand), but in practice no progressive loading actually happens. All matched skills have their full content injected into the system prompt in a single step.

Current Behavior

When agent.run(task) is called:

1. handle_skills(task) is invoked
2. DynamicSkillsLoader.load_relevant_skills(task) computes cosine similarity and filters skills
3. All matching skills' full content is immediately concatenated into system_prompt

The load_full_skill(skill_name) method exists as a public API but is never called internally — it's an orphaned method with no integration into the agent's execution pipeline.

Expected Behavior (True Progressive Loading)

A proper tiered implementation would look like:

• Tier 1 (initialization): Only inject skill name + description into the system prompt, keeping it lightweight
• Tier 2 (runtime, on-demand): When the agent determines a skill is needed during conversation, dynamically load the full SKILL.md content via load_full_skill() and append it to the context

This requires a mechanism for the agent to "request" a skill at runtime (e.g., a tool/function call), which is currently missing.

Why This Matters

• Token waste: Loading all full skill content upfront bloats the system prompt, especially with many skills
• Context window pressure: For agents with many skills, this can consume significant context unnecessarily
• Documentation mismatch: The docs and code comments reference "Tier 2 progressive disclosure" but the implementation doesn't deliver it

Relevant Code

• swarms/structs/dynamic_skills_loader.py — DynamicSkillsLoader class
• swarms/structs/agent.py — handle_skills(), _load_dynamic_skills_for_task(), load_full_skill()

Suggested Approach

1. In Tier 1 mode, only append a skill summary (name + description) to the system prompt
2. Expose load_full_skill as an agent tool/function that the LLM can call mid-conversation
3. When the LLM invokes the tool, load and inject the full skill content into the context dynamically

[kyegomez]

@pengjunfeng11 thanks for pointing this out to me, it will be fixed