Warp builds self-improving agents on Claude

Learn really Warp devised a simple improvement shape that anyone tin usage to create self-improving agents.

In our series, , we item really startups are transforming their industries pinch AI. In this article, we stock really Warp turned stateless personification feedback into a self-improvement loop for its agents.

Agents request to grip recurring tasks reliably and effectively. A first-pass punctual that gets 80% of the task correct tin create a noisy and annoying acquisition for the user. Warp learned this the difficult way, and utilized this to pass its merchandise strategy, creating an improved acquisition for astir 1M developers worldwide.

Warp, the AI-powered terminal and agentic improvement environment, builds connected the Claude Platform. The squad ran into this “noisy experience” problem pinch their soul codification reappraisal agent. Engineers complained that their supplier made unhelpful comments and produced low-quality output.

The squad initially tried stopgap solutions, for illustration manually rewriting the punctual based connected observed codification reappraisal failures. This made output much usable but didn’t scale. Improving discourse files for illustration AGENTS.md besides helped, but was acold from a complete fix. 

Ultimately, they realized, the existent rumor was that feedback to an agent, nary matter what its purpose, typically disappears erstwhile the convention ends, removing captious discourse from the agentic loop. Their solution: an Agent Skills-based model to create self-improving agents wherever feedback compounds complete clip to continually refine and heighten supplier output. 

Read connected to study really they built it pinch skills connected apical of the Claude Platform. 

Agent self-improvement loops built connected skills

The cardinal method is simply a self-improvement loop utilizing skills, which are record based encodings of knowledge that support instructions retired of the earthy prompt. Warp evolved a self-improving supplier architecture consisting of 2 skills, pinch quality feedback successful between. 

The inner/base skill holds the functional domain knowledge and instructions. For example, erstwhile a PR is opened, Warp’s codification supplier executes utilizing that guidelines accomplishment and discourse to nutrient its review.

Human feedback on supplier output is simply a captious constituent for the self-improvement loop. For codification reappraisal this could beryllium thing arsenic elemental arsenic a thumbs up, but the much definitive the better. 

 “A quality could affirm, ‘this was a good, useful comment’,” Warp laminitis Zach Lloyd explains, “But the quality could besides springiness elaborate reasons why a codification reappraisal wasn't good. Specifics for illustration ‘you suggested renaming this variable, but our codification guidelines normal is this type of world adaptable uses this peculiar naming context’ show the supplier really to do it correct adjacent time.”

The outer/improver skill functions arsenic an perceiver supplier that runs connected a schedule alternatively than per-task. It pulls the accumulated quality feedback, compares what the supplier suggested against really humans responded, and proposes a small, focused edit to the guidelines skill.

Because skills are plain files, agents are highly bully astatine updating them. These updates, which are reviewable, approvable, and mergeable, tin travel done a normal PR/code-review workflow; erstwhile merged, the adjacent tally of the soul accomplishment inherits the improvement. 

Warp now runs this shape crossed its full open-source repo, pinch abstracted spec-writing, review, and triage agents, each carrying their ain self-improvement loop.

“File-based skills are a measurement of encoding knowledge for agents without putting that knowledge straight successful the prompt, arsenic thing the supplier tin simply look up successful the people of doing its job,” says Zach. “The model is really elemental actually: there's the guidelines domain-specific accomplishment and past there's the improver accomplishment that refines  that domain-specific skill. This simplicity is the beauty of this approach.”

How to constitute self-improving skills for agents

Here are immoderate of the Warp team’s tried and existent tips for penning self-improving skills for agentic loops: 

• Write principles, not rules. "Construct the accomplishment arsenic though you're instructing a smart person, not for illustration you're programming a computer,” Zach says. “Including guidance successful the accomplishment for illustration ’Look for repeated code’ provides amended guidance than exhaustive adaptable naming rules.” 
• Explain the why. Providing the rationale down the norm lets the supplier logic astir the problem alternatively of pursuing rigid instructions, again allowing for amended generalization. 
• Make feedback effortless to give. Capture it wherever group already work, for illustration by commenting straight connected a PR aliases issue. Also, make this hap automatically, pinch nary other submission step. “Low clash is what keeps awesome flowing,” Zach notes. “If you make it excessively difficult you're not going to get the feedback and you're not going to beryllium capable to amended the skill."
• Keep skills mini and usage progressive disclosure. A bully skill record isn't large; it references assets files and scripts alternatively than dumping everything into discourse astatine once. 
• Feedback value > volume, but measurement helps. A mini magnitude of detailed, domain-specific feedback from a elder technologist tin beryllium worthy much than tons of cursory feedback because binary thumbs up/down doesn't opportunity why. “You tin get really bully awesome moreover from a comparatively mini sample size if it's very elaborate feedback from a personification astir domain circumstantial knowledge that the supplier different would person nary measurement of getting,” Zach continues. “That said, the bigger the corpus of value signal, the better. At Warp we're utilizing a loop to negociate our full unfastened root repo. We person hundreds of group contributing and we're doing thousands of codification reviews.”
• Put other effort into the improver skill. Putting other effort into penning the improver accomplishment (the perceiver agent) pays disconnected beyond the contiguous supplier loop, because improver skills are very reusable crossed different usage cases.  “Outside of the domain circumstantial knowledge component, this is simply a reasonably reusable mechanism—the improver accomplishment for a codification reappraisal supplier is not that different from the improver accomplishment for immoderate different agent.”

The loop successful action: Warp’s rumor triage agent 

Warp’s rumor triage agent demonstrates the self-improving supplier skills framework. The shape is triggered whenever personification files a caller GitHub issue: a GitHub Action fires an supplier that analyzes the rumor for complexity and feasibility, assigns labels, and suggests a guidance for the fix. That triage supplier runs disconnected an soul accomplishment record holding the domain knowledge astir what each explanation intends and really to investigation the codebase earlier acting.

On a sample issue, the first-stage soul accomplishment did a coagulated occupation but missed 1 label, fresh to spec, which signals that a contributor tin commencement building merchandise and method specs against the issue. A maintainer connected the Warp squad caught the spread and near feedback straight connected the issue, precisely wherever the activity was happening. Critically, he explained some what he expected and why he expected it: actionable feedback easy for the supplier to sorb later. 

The outer improver accomplishment runs successful Oz, Warp's supplier orchestration platform, arsenic a scheduled “update triage” agent. The supplier authenticated to GitHub, ran a Python book bundled pinch the accomplishment to propulsion caller issues carrying feedback, summarized them into a JSON file, and publication that backmost into context. The bundled book is itself a champion practice; skills tin reference assets files alternatively of penning caller codification connected each run.

From there, the supplier identified the actual feedback signals successful the maintainer comments and projected the smallest edit that captured them. It opened a PR editing the soul accomplishment to use the "ready to spec" explanation erstwhile an rumor describes a existent problem, moreover though the nonstop UI aliases UX style is not yet defined. 

Because the full update is simply a accomplishment file, it moves done the normal code-review workflow. The PR arrived pinch a explanation explaining which signals prompted the alteration and what it altered. A quality reviews, approves, and merges, and the adjacent tally of the triage accomplishment inherits the caller knowledge. That last quality measurement closes the loop and keeps a personification successful power of what really changes.

This is the aforesaid system Warp now runs astatine standard crossed its open-source repo, wherever spec-writing agents, reappraisal agents, and triage agents each transportation their ain self-improvement loop. 

Any agent, nary matter what its task, gets amended complete clip if you build 1 of these loops into it from the commencement to seizure quality feedback signals, move them into accomplishment updates, and grow agents from one-off helpers into tin systems that compound crossed your org.

View the afloat webinar for a unrecorded demo and deeper chat of really Warp uses Claude to build agents that study from squad feedback and amended themselves complete time.

Start building pinch the Claude Platform today.