Warp, the AI-powered terminal used by 800,000 monthly developers, recently detailed how they build autonomous, self-improving agents on the Claude Platform.
According to Anthropic's recent case study, Warp ran into a classic automation wall: a first-pass prompt that gets a task 80% right creates a noisy, frustrating experience for the user. For Warp, this manifested in an internal code review agent that routinely made unhelpful comments. Stopgap fixes, like manually rewriting prompts after observing failures, didn't scale.
The core problem was statelessness. Whenever a session ended, the user's feedback disappeared, removing critical context from the agent's next run. Warp's CEO, Zach Lloyd, explained their solution: an Agent Skills-based framework that captures user feedback to continually refine the agent's output over time.
The Dual-Agent Improvement Loop
Warp's architecture relies on two distinct skills—file-based encodings of knowledge that keep heavy instructions out of the raw prompt.
- The Inner/Base Skill: This holds the functional domain knowledge. When a task is triggered, the agent executes using this base skill to produce its output.
- The Outer/Improver Skill: This is an observer agent that runs on a schedule rather than per-task. It pulls accumulated human feedback, compares the agent's initial suggestion against how the human responded, and proposes a small, focused edit to the base skill file.
Because these skills are plain text files, the updates flow through a standard review workflow. A human approves the update, and the next run of the base skill inherits the improvement. Warp now runs this pattern across its open-source repo, supporting a massive footprint of over 40 million total Warp Agent conversations to date.
Lloyd offers a few practical rules for this framework: write principles instead of exhaustive rules, explain the why behind a task, and prioritize highly detailed feedback over binary thumbs-up/down signals. Crucially, feedback must be effortless to give—captured exactly where the user is already working to maintain a strong signal.
What This Means for SMB Operations
While Warp is a high-growth developer environment, its self-improving architecture is a direct blueprint for small and mid-sized businesses automating their administrative and operational tasks.
Most SMBs hit the exact same "80% accuracy" plateau when deploying AI. You might build an agent to draft customer service replies, categorize inbound leads, or process vendor invoices. Initially, it saves a tremendous amount of time. But if it stalls at 80% accuracy, human reviewers end up spending as much time fixing the remaining 20% of errors as they would doing the work from scratch. Because the agent is stateless, the human corrects the invoice directly in the accounting software, but the agent learns nothing. The next day, it makes the exact same mistake.
Breaking past this plateau requires shifting from a static prompt to a self-improving loop. You do not need a massive engineering team to replicate Warp's concept. You just need a systematic way to manage your agent's instructions.
Building Your Own Feedback Loop
To apply Warp's approach to SMB operations, treat your agent's "skills" as your standard operating procedure (SOP) documents.
Separate instructions from the prompt. Keep the initial prompt simple, but point the agent to an external SOP file to reference for its rules. If you dump every edge case into your initial prompt, it becomes unmanageable.
Capture the specific correction. If your dispatching agent assigns a commercial plumbing job to a residential technician, don't just manually reassign the ticket in your CRM. Record the specific rule: "Technician A only handles residential properties; commercial jobs require Technician B." As Lloyd notes, the detailed why matters more than a simple "wrong assignment" flag.
Automate the SOP update. You can set up an "improver" automation that runs every Friday. Have it review all the manual corrections logged during the week, compare them against the current SOP document, and draft an updated set of rules. An operations manager can review this proposed update—just like a code review—and approve it. By Monday, your dispatch agent is operating on the improved rules.
Agentic automation is not about writing the perfect prompt on day one. It is about building a low-friction system that gets incrementally better every time it makes a mistake. For SMBs, adopting this continuous improvement model is the difference between an automation that requires constant babysitting and one that sustainably scales your business.