How Warp Builds Self-Improving Agents on Claude with Skills!

Hi, it's Shiichan! Today I want to share a clever trick from Anthropic's Claude Blog about the AI terminal tool Warp.

Claude Blog claude.com

What was announced?

Anthropic's Claude Blog published an article explaining how Warp builds "self-improving agents" on Claude. Warp is an AI terminal and agentic development environment founded in 2020, led by CEO Zach Lloyd. The company has raised 73 million dollars, serves 800,000 monthly developers, and is used at 56 percent of Fortune 500 companies. The post walks through how Warp's engineering team uses Agent Skills, Claude's skills feature, to help agents grow over time.

The story so far

Warp's team found that their early agent prompts hit around 80 percent accuracy, yet the user experience still fell short. The reason: as the article puts it, "feedback to an agent, no matter what its purpose, typically disappears when the session ends." That meant the same mistakes kept resurfacing session after session.

What changes

To fix this, Warp built a self-improving framework on top of Agent Skills. The core idea is giving each agent two kinds of skills.

• Inner (base) skill: holds the domain knowledge and instructions for the actual task, such as analyzing a code review
• Outer (improver) skill: an observer agent that runs on a schedule, analyzes accumulated human feedback, and proposes targeted edits to the base skill

As Warp founder Zach Lloyd explains, "File-based skills are a way of encoding knowledge for agents without putting that knowledge directly in the prompt, as something the agent can simply look up in the course of doing its job."

Thanks to this setup, feedback no longer vanishes when a session ends — it accumulates in the base skill, so the agent shows up smarter the next time it runs. Warp now reports over 10 million Claude Code sessions run inside Warp (400,000-plus weekly), and 40 million total Warp Agent conversations. Several agents, covering spec writing, review, and triage, run on this same pattern.

Dive Deep

The article walks through a GitHub issue triage agent as a concrete example.

• It triggers via GitHub Actions whenever a new GitHub issue is filed
• It analyzes complexity and feasibility, then assigns labels automatically
• The improver skill runs on a schedule through Oz, Warp's internal agent orchestration platform
• Accumulated feedback flows through as a pull request, following the normal code-review workflow
• The next triage run inherits whatever improvements were merged

The stack behind this includes Rust, Golang, and GitHub Actions, orchestrated through Oz, and built on top of the Claude Platform's Skills API.

The article also lists a set of best practices for writing self-improving skills.

• Write principles, not rules: instruct the agent the way you'd brief a smart person, not the way you'd program a machine
• Explain the rationale: give the agent room to reason instead of following rigid instructions
• Make feedback effortless: capture it where people already work, so friction doesn't choke off the signal
• Keep skills small: reference resource files instead of dumping everything into context
• Prioritize feedback quality: input from domain experts matters more than a high volume of superficial feedback
• Invest in improver skills: they are highly reusable across different use cases

With this approach, the article argues, agents stop being one-off helpers and turn into capable systems that compound across an organization.

Wrap-up

• Anthropic's Claude Blog featured how Warp uses Claude's Agent Skills to build self-improving agents.
• The problem: feedback used to disappear once a session ended. Warp solved it with a two-skill pattern of a base skill plus an improver skill.
• Warp now counts over 10 million Claude Code sessions and more than 40 million total Warp Agent conversations.
• Using a GitHub issue triage agent as an example, the article shows how GitHub Actions, the internal Oz orchestration platform, and code review all tie together.
• Worth reading for engineers and platform teams running their own agents who are wondering how to make feedback actually stick.