The most interesting thing an agent can do is get better at its job without you reprogramming it. A self-improving agent does exactly that: when it solves a non-obvious problem, it captures the solution as a new, reusable skill. The next time a similar task appears, it doesn't re-derive the answer — it already has one. Your agent's capabilities start to mirror the work you actually do.
It's no surprise this is one of the most reached-for categories. On OpenClaw, the Self-Improving Agent skill is the headline option.
Self-improvement on OpenClaw — install the loop
On OpenClaw, self-improvement is something you install. The flagship is the Self-Improving Agent skill, with strong alternatives like agent-evolver ("self-evolution engine enabling agents to learn from experience") and agent-reflect (self-improvement through conversation analysis):
openclaw skills install self-improving-agentBecause it's modular, you can choose how aggressive the loop is and combine it with a memory skill so the lessons persist. The flip side: you're explicitly opting your agent into rewriting its own behavior, so this is the skill that most rewards pairing with vetting.
Self-improvement on Hermes — the native loop
Hermes is built around this idea rather than bolting it on. It's described as "the only agent with a built-in learning loop" — it creates skills from experience, improves them during use, and maintains its own skill library through an autonomous Curator that grades, consolidates, and prunes on a regular cycle. There's nothing to install; switching it on is switching on Hermes.
The "Claudeception" pattern from the Hermes community sharpens this further: when the agent cracks a weird problem, it writes that knowledge back as a skill automatically, so the loop tightens with every hard task.
OpenClaw vs Hermes at a glance
- OpenClaw — install the Self-Improving Agent skill (or an alternative) and tune the loop yourself.
- Hermes — native learning loop plus a self-pruning Curator, on by default.
Running it from your phone
A self-improving agent changes its own behavior, so you'll want to watch what it generates. Onepilot lets you review and approve new skills over SSH from an iPhone before they run with real access — and pairs naturally with the skill-security page. Set it up on OpenClaw or Hermes and keep the approval step in your pocket.