Warp builds self-improving agents on Claude

Warp, the AI-powered terminal and agentic development environment built on the Claude Platform (stack: Rust, Golang, GitHub Actions, internal agent orchestration platform Oz), found that its internal code review agent produced a 'noisy experience' when a first-pass prompt got only 80% of the task correct—engineers complained about unhelpful comments and low-quality output. Initial workarounds like manually rewriting prompts or improving AGENTS.md context files helped but didn't scale. The team realized the real issue: feedback to an agent typically disappears when the session ends, removing critical context from the agentic loop.

To fix this, Warp built a self-improving agent architecture using Agent Skills, file-based encodings of knowledge that keep instructions out of the raw prompt. The system uses two skills with human feedback in between: an inner/base skill containing functional domain knowledge and task instructions (executed when a PR is opened), and an outer/improver skill that acts as an observer agent running on a schedule. Human feedback—from a simple thumbs-up to detailed reasons like 'you suggested renaming this variable, but our code base convention is this type of global variable uses this particular naming context'—tells the agent how to do it right next time.

Warp has raised $73M and reports 800K monthly developers, 56% Fortune 500 adoption, 10M Claude Code sessions inside Warp to date (400K+ per week), and 40M total Warp Agent conversations. The company says this framework creates a self-improvement loop where feedback compounds over time, refining agent output for nearly 1M developers worldwide.