Warp Factories: Automated Skill Improvement via Built-in Feedback Loops

X postZach LloydAnnouncement

Zach Lloyd shares that Warp's software factory has automatically generated and merged ten skill improvement PRs. This workflow is powered by Warp Factories' new self-improvement feature, which monitors agent runs for inefficiencies and autonomously proposes fixes to agent prompts and skill configurations.

Key points

  1. Warp Factories automates the software development lifecycle using dedicated agents for triage, specifications, implementation, and code review.
  2. Configurable 'Scorers' evaluate a sampled subset of agent runs against criteria like efficiency, compliance, and code quality using defined thresholds.
  3. A dedicated self-improvement agent periodically inspects failing runs, analyses execution transcripts, and identifies root causes.
  4. Proposed fixes to skills and configurations are delivered as pull requests or review branches for human approval.
  5. In one cited instance, the system resolved an agent loop executing 275 screenshot calls during computer use by updating the skill to prefer browser DevTools for zooming.

Why it matters

Doom-loops and repetitive tool calls can rapidly escalate token consumption and operational costs in autonomous agent workflows. Automating the detection, diagnosis, and patching of agent skills removes the friction of manual prompt tuning and treats agent behaviour as continuously tested code.

Analysis

Scaling coding agents requires robust mechanisms to catch edge cases that manual reviews easily miss. Warp's approach structures evaluation into discrete scorers and feeds execution traces directly into a secondary agent dedicated to skill maintenance. Generating human-reviewable PRs keeps the deployment pipeline grounded while allowing the agentic factory to adapt autonomously over time.

Original post

Zach Lloyd

@zachlloydtweets · 1 September 2026

We're now at 10 skill improvement PRs merged, automatically made by our factory. Ex: This system caught an agent using 275 tool calls in a single run to screenshot its work during computer use. It patched the skill to prefer browser DevTools for zooming; No runaway costs since t.co/DdBPEHTtDP

172 likes10 replies298 bookmarks24.3k views

View on X
Selected replies (5)
  • @OMID_0909 @zachlloydtweets Automatically learning from agent runs is powerful. EKOS adds an evidence-backed knowledge layer to make that operational context reusable across future runs. 🤝🫂 t.co/UX2XsqcGOe
  • @FlorinIluta @zachlloydtweets 275 tool calls just to screenshot its work is wild :)) Having the system catch that and turn it into a skill fix is pretty damn cool.
  • @AIQuanting @zachlloydtweets the 275 calls one is easy, it screams in the bill. more curious about the foreman one, disclosing unresolved tracker failures. what surfaced that? theres no cost spike when an agent just quietly doesnt mention something
  • @mktpavlenko @zachlloydtweets 275 tool calls for screenshots is how a tiny inspection habit turns into a bill nobody notices until the factory scales.
  • @OMID_0909 @zachlloydtweets t.co/UxnEsQ09wR

Mentioned links

Links and tools from this entry

Summary generated by AI (Gemini 3.8 Flash, 27 September 2026). It may contain errors – the original source is authoritative.

Related entries

Similar in content, found by the AI search.

  • X post:Ben Holmes

    Building a Software Factory: Zach Lloyd's Crawl-Walk-Run Framework

    Ben Holmes highlights an architectural guide by Zach Lloyd detailing how engineering organisations can progressively adopt an agentic software factory model, moving from isolated point automations to governed cloud workflows.

    1.3kbookmarks124.2kviews

    AI & LLMs· Resource

  • Link:Abacus.AI

    Abacus.AI AutoBots: Self-Improving Autonomous Agents

    Abacus.AI has launched AutoBots, a platform offering recursively self-improving autonomous agents designed to handle end-to-end workflows. Configured via text prompts, the agents run continuously without human intervention and refine their strategies across successive executions.

    AI & LLMs· Tool

  • Video

    Video:CURT

    Harness Engineering: How to Build a Software Factory with AI Agents

    In his talk at the AI Engineer World's Fair, Tessl's Dru Knox introduces the concept of a software factory, where coding agents produce all user-facing code whilst engineers build and maintain the factory infrastructure. He outlines harness engineering (also referred to as loop engineering) as the guiding discipline, dividing engineering work into inner, outer, and meta loops to improve autonomy, automation, and overall code quality.

    AI & LLMs· Talk

Back to the knowledge base

Let's talk about your project

Get in touch

Locations

  • Mattersburg
    Johann Nepomuk Bergerstraße 7/2/14
    7210 Mattersburg, Austria
  • Vienna
    Ungargasse 64-66/3/404
    1030 Wien, Austria

Parts of this content were created with the assistance of AI.