Equipping Agents for the Real World with Agent Skills
Source: https://www.anthropic.com/research/agent-skills Author: Anthropic Date: 2025-10-18
Summary
Anthropic’s post on “agent skills” — reusable, composable capabilities that agents can acquire and use. Frames agent skill acquisition as a distinct challenge from model capability.
Key Claims
- Skills vs. capabilities: skills are discrete, reusable actions (search web, write file, call API); capabilities are general (reasoning, understanding). Both matter but skills are acquired differently.
- Skill acquisition: agents learn skills through: (1) pre-built skill libraries, (2) few-shot demonstration, (3) RL on task outcomes.
- Composability: complex agent tasks decompose into skill sequences. An agent that knows individual skills can combine them for novel tasks.
- Reliability: skills need to be reliable — a skill that works 95% of the time creates cascading failures in multi-step tasks (95%^10 = 60%).
- The open problem: skill generalization — how to acquire skills from one domain that transfer to another.
Entities
Concepts
- Coding Agents — skill-based coding agent architecture
- RL Infrastructure — RL for skill acquisition
- Context Engineering — skill descriptions in context