Sharpagent Self Evolving
ReviewAudited by ClawScan on May 11, 2026.
Overview
This skill does not install code, but it asks the agent to automatically learn from every task and save lessons from task logs without clear limits or review controls.
Install this only if you intentionally want a persistent self-improvement workflow. Before using it, decide where learning records will be stored, how to remove them, whether sensitive tasks should be excluded, and whether the reflection loop should require explicit approval after each task.
Findings (3)
Artifact-based informational review of SKILL.md, metadata, install specs, static scan signals, and capability signals. ClawScan does not execute the skill or run runtime probes.
The agent may spend time on self-improvement or continue working after the user expected the task to be finished.
This changes normal task completion into a mandatory follow-on workflow and can cause the agent to continue reflecting or experimenting even when the user only requested the original task.
After every task, do a structured reflection. ... Every task completion (mandatory) ... Major errors mid-task (force deep mode)
Make the loop explicitly opt-in per task, require user confirmation before experiments, and provide a clear stop or skip instruction.
Private task details or incorrect lessons could be saved and later influence unrelated work.
The skill is meant to read prior task context and logs, then store lessons for future use, but the artifact does not define where this memory is kept, how long it persists, how sensitive data is excluded, or how users can review/delete it.
reads: - Task - LearningEntry ... writes: - LearningEntry - ImprovementHypothesis ... "Access to read task output and logs"
Define a scoped storage location, retention policy, secret redaction rules, project/user boundaries, and a review/delete process for learning records.
Users may over-trust the skill's safety or provenance based on a self-declared status.
The skill self-attests that it is verified, while the supplied registry context lists the source as unknown and provides no homepage or independent verification evidence.
trust_level: verified
Treat the verification claim as unproven unless the registry or publisher provides independent provenance and review evidence.