Skill Library

Persist newly composed skills that proved useful so later tasks can reuse them as first-class tools instead of rediscovering the composition.

When an agent frequently chains the same tools, those transitions are candidates for composition into a larger composite tool. Voyager-style systems go further with a skill manager: newly created skills (often programs) that helped complete a task are added to a skill library — conceptually an evolving extension of the tool inventory — for reuse on future tasks. Because such skills are agent-authored, they can quietly drift out of date or out of policy (self-authored skill drift).

This is external memory for capabilities, not just facts: the harness grows the action repertoire from successful trajectories. Pair with tool selection ablation so the library stays useful rather than accumulating unused or unreliable skills.

Close relatives store other reusable units distilled from trajectories: induced workflows (Agent Workflow Memory), cached plan templates (agentic plan caching), or itemized strategy bullets (ACE). See prompt evolution vs playbook accumulation for where these sit among context-adaptation designs.

---

Sources

• AI Engineering (Chip Huyen), ch. 6
• ACE (Zhang et al.), App. B.1, §3.1