Skills: Packaged Procedures with Progressive Disclosure

Give an agent a written procedure for a task, prepared by people who know the task well, and on average it does better. A benchmark called SkillsBench gave agents such procedures for 87 tasks, and the average pass rate rose from about 34 percent to about 50 percent.

Yet on 13 of those tasks, the agent did worse with the procedure, and agents that wrote their own procedures scored below agents that had none. Both results depend on how a procedure enters the context and which of the agent's own approaches it replaces.