xskill-self-improving-object

XSkill Self-Improving Graph Objects

Overview

The improvement mechanisms earlier in Ch7 (prompt refinement, SEAL data generation, fine-tuning) all modify the agent itself. Knowledge augmentation is lighter: it accumulates knowledge from past executions and retrieves it at inference time, touching neither the model nor its prompts. The motivating measurement from the chapter: on the Kaggle GameArena chess benchmark, 78% of Gemini-2.5-Flash losses were illegal moves, rule violations rather than weak strategy. The agent kept repeating the same category of mistake because it had no memory of past failures.

XSkill (Jiang et al., 2026) extracts two complementary knowledge types from trajectories:

• Experiences operate at the action level. Each execution node's input, action, and outcome becomes a candidate experience record. Experiences alone reduce tool errors from 29.9% to 16.3% (a 45% reduction).
• Skills operate at the task level. The path from the root query node to a successful resolution node becomes a candidate multistep skill. Together with experiences, the average success rate rises from 33.6% to 40.3%.

Cognee closes the remaining gap: XSkill's skills are static artifacts. Cognee treats a skill as a first-class graph node (the SkillNode example) with execution records, a success rate, and an amendment history. Its four-stage pipeline is add (parse SKILL.md, compute content hash) then cognify (extract trigger phrases and complexity) then search (route by which skill SUCCEEDS at similar tasks) then learn (log an observation per execution). When a skill degrades, amendify() rewrites it against the last 10 failures, validates the amendment against held-out records, and rolls back on failure.

When to Use

• After an agent has accumulated execution traces and you want it to stop repeating avoidable mistakes without retraining
• Environments that drift (a Kubernetes API change, a new CI/CD stage) where a static SKILL.md silently goes stale
• Routing among several overlapping skills where description similarity picks the wrong one and demonstrated success should decide
• Building the self-evolution loop on top of the execution-graph substrate