Agent skills deliver real value in production systems, but most teams are implementing them incorrectly, according to new research. Companies like Atlassian, Canva, and Figma have successfully integrated agent skills into their products. Atlassian's Rovo uses them to automatically triage Jira tickets, draft Confluence pages, and route service requests without manual prompts. Canva and Figma leverage similar capabilities with Claude.
However, the research identifies a critical gap between how teams build these systems and how they should work. The core issue centers on execution patterns. Most teams treat agent skills as simple function calls with minimal context or error handling. This approach fails when tasks require multi-step reasoning, fallback logic, or dynamic decision-making.
Effective agent skill implementation requires several shifts. First, teams need explicit task decomposition. Rather than giving an agent a single broad skill, breaking tasks into smaller, verifiable components improves reliability. Second, error recovery matters more than perfect execution. Systems should gracefully degrade and offer alternatives when primary paths fail.
Third, context matters enormously. Agents perform better when they have access to relevant system state, user preferences, and conversation history. Teams building skills with minimal context essentially handicap their systems.
The research also highlights validation gaps. Most teams skip rigorous testing of agent behavior across edge cases. Production failures spike because teams assume agent skills work like traditional APIs. They don't. Agent behavior emerges from interaction between the model, the skill definition, and the execution environment.
Teams at Atlassian, Canva, and Figma succeed because they treat skill implementation as an iterative process. They monitor agent decisions, collect failure patterns, and refine skill definitions based on real-world performance. This contrasts sharply with one-shot implementations many organizations attempt.
The broader lesson: agent skills are viable, but their success depends on