What caught my attention is not the “self-improving agents” label. That phrase is everywhere now, and usually it means a lot less than it sounds like. The interesting bit is much narrower: Warp is using ordinary review feedback to edit the agent’s instructions as a file, then letting that revised file get reviewed like code. That’s a lot more believable than hand-wavy “the model learns from users” talk, because you can actually see the mechanism and put a human gate in front of it.
I also like that they’re not pretending feedback is magic. The article is pretty honest about the first version being noisy, the prompt tweaks helping but not scaling, and AGENTS.md-style context being incomplete. That matches what most people discover: you can spend forever sanding the prompt and still end up with an agent that is technically competent but socially annoying. The problem isn’t just accuracy. It’s that the agent keeps forgetting the local norms that matter to the team.
The file-based skill idea makes sense to me because it treats knowledge as something you can version, diff, and review. That’s a much cleaner story than stuffing more and more into the prompt. But I do think there’s a subtle risk here that the article only gestures at: if the improver keeps making “small edits,” you can still end up with a pile of local patches that encode yesterday’s preferences instead of a durable policy. The human review step helps, but it doesn’t automatically solve drift. You still need someone who notices when the skill has become overfit to one weird issue thread.
The other thing I’d be cautious about is the claim that this pattern is broadly reusable. Maybe it is, but I’d want to see it outside code review and issue triage before I got excited. Developers are unusually good at leaving structured feedback and unusually tolerant of workflows that route through PRs. A support agent, a sales agent, or an internal ops agent might not generate feedback that is nearly as clean. The loop could degrade fast if the comments are vague, political, or contradictory.
Still, this is the first “self-improving agents” pitch in a while that doesn’t feel like pure marketing vapor. It’s basically: capture criticism where work already happens, distill it into a small instruction change, and require a person to approve that change. That’s not glamorous, but it is implementable. And in this space, implementable is rare enough to be interesting.
Reference: How Warp builds self-improving agents on Claude | Claude by Anthropic
