"Agent Skills" feature

[genis-lainez-moreno]

Summary

We should explore adding a "Skills" feature to smolagents, inspired by Anthropic's Agent Skills standard. While tools provide agents with "hands" to execute actions, Skills provide the "knowledge" or "onboarding" of how and when to perform complex, domain-specific workflows.

Background

Currently, smolagents excels at CodeAgent and ToolCallingAgent workflows where tools are explicitly defined. However, as agents grow more complex, saturating the system prompt with every possible tool description becomes token-expensive and can degrade performance.

[KeepALifeUS]

This is a great idea! I've been thinking about similar patterns.

The tools vs skills distinction is key:

• Tools = atomic actions (API calls, file operations)
• Skills = composed workflows (multi-step procedures with decision points)

One approach I've found effective:

Instead of saturating the system prompt, you can implement skills as shared state patterns. The agent doesn't need to know every skill upfront — it reads the current state and knows what's expected.

# Skill as state schema
skills = {
    "code_review": {
        "steps": ["read_file", "analyze", "suggest_fixes"],
        "current_step": 0,
        "context": {}
    }
}

Benefits:

1. Token efficiency — Agent only sees the current step context, not the entire skill definition
2. Resumability — If interrupted, just resume from saved state
3. Composability — Skills can invoke other skills by updating state

How this relates to Anthropic's approach:

Anthropic's C compiler project used a similar idea — file-based coordination where agents don't directly communicate but read/write to shared state. The "skill" emerges from the state transitions rather than being encoded in prompts.

Working example with 4 specialized agents: https://github.com/KeepALifeUS/autonomous-agents

The key insight: skills can be emergent from coordination patterns rather than explicitly defined. This scales better as complexity grows.

Would love to see smolagents explore this direction!