Warp, the AI-powered terminal and development environment, has built a self-improving agent architecture on the Claude Platform to solve recurring task reliability issues. Founded in 2020 by Zach Lloyd, Warp faces the common problem where a first-pass prompt achieving 80% task correctness creates noisy user experiences. The team initially tried manual prompt rewriting and improved context files, but found these solutions did not scale.

The core issue was that feedback to agents typically vanishes when sessions end, removing critical context from the agentic loop. Warp's solution is an Agent Skills-based framework where feedback compounds over time to refine agent output. The architecture consists of two skill types: an inner/base skill holding functional domain knowledge and instructions, and an outer/improver skill acting as an observer agent.

The improver runs on a schedule, pulling accumulated human feedback and comparing agent suggestions against human responses to propose focused edits to the base skill. Because skills are plain files, updates are reviewable, approvable, and mergeable through normal PR workflows. Once merged, the next run of the inner skill inherits the improvement.

Warp now runs this pattern across its entire open-source repo with separate spec-writing, review, and triage agents, each carrying their own self-improvement loop. The framework allows compounding improvements without polluting raw prompts, creating a sustainable path toward increasingly reliable AI-assisted development.