Skill Learning: Bringing Continual Learning to CLI Agents Fresh

Published: December 2, 2025 | Original Post

---

Dynamic skill learning through experience that improves agents over time rather than degrading them.

How Skill Learning Works

A two-stage process:

1. Reflection Stage

The agent evaluates a task trajectory to:

• Determine if the task was successfully completed
• Assess reasoning validity
• Identify repetitive patterns suitable for abstraction

2. Creation Stage

A learning agent generates skills containing:

• Potential approaches to tasks
• Common pitfalls to avoid
• Verification strategies

Performance Results (Terminal-Bench 2.0)

Method	Improvement	Notes
Trajectory-only skills	21.1% relative	9% absolute improvement
Trajectory + feedback	36.8% relative	15.7% absolute improvement
Cost reduction	15.7%	10.4% fewer tool calls

Enriching reflections with textual error feedback produces higher-quality, more robust skills by encoding failure modes alongside successes.

Integration with Agent Memory

Skills complement existing memory in a hierarchical structure:

• Core Memory / System Prompt -- Evolving prompts specific to individual agent state, applied across all tasks
• Skills / Filesystem -- Task-specific, modular files designed for interchangeability between agents

Usage

Access Skill Learning through Letta Code by calling the /skill command after agent interactions.