Anthropic examines progress towards recursive self-improvement in AI

LinkAnthropicResearch

The Anthropic Institute shares internal metrics and external benchmark trends demonstrating how AI agents are accelerating the development of subsequent AI models. As of May 2026, over 80% of merged code at Anthropic is written by Claude, with engineers merging 8x more code per day than in 2024. While full recursive self-improvement remains constrained by models' lack of strategic judgement in goal selection, progress is advancing rapidly.

Key points

  1. Over 80% of code merged into Anthropic's codebase as of May 2026 was authored by Claude, up from low single digits before Claude Code launched in February 2025.
  2. Engineers at Anthropic merged 8x as many lines of code per day in Q2 2026 compared to 2024 rates.
  3. The duration of tasks models can reliably handle independently is doubling roughly every four months, progressing from 4-minute tasks (Claude Opus 3) to 1.5 hours (Claude Sonnet 3.7) and 12 hours (Claude Opus 4.6).
  4. Claude Mythos Preview demonstrated task endurance of at least 16 hours in METR evaluations, with median internal survey respondents reporting a 4x output increase.
  5. On open-ended, underspecified tasks, Claude's success rate without human intervention reached 76% in May 2026.
  6. An automated Claude-based code reviewer now checks pull requests, with retrospective testing showing it would have caught about one third of past production bugs on claude.ai.

Why it matters

Delegating the creation of AI systems to AI itself compounds development velocity and brings recursive self-improvement closer. This shift heightens security and alignment risks, requiring far more rigorous internal oversight, automated code review, and behavioral containment.

Analysis

Anthropic's report provides concrete internal operational data illustrating how software engineering at leading AI labs has transformed. Rather than serving as inline auto-completers, agents like Claude Code operate over extended time horizons, diagnosing cluster outages and refactoring large codebases. The saturation of software and replication benchmarks (SWE-bench, CORE-Bench) confirms this trajectory externally. The missing link for closed-loop recursive self-improvement remains senior-level judgement—deciding what experiments to prioritize and which architectural problems matter most.

Mentioned links

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.

  • Article:Addy Osmani, Morgan Westlee Lunt, Qing Wang

    Code Review at Anthropic: How the Claude Code Team Reviews AI-Generated Code

    The Claude Code engineering team explains how they handle code reviews when agents write around 90 percent of the codebase. Instead of exhaustive manual line-by-line checks, they rely on pre-PR verification, automated multi-agent reviews, risk-tiered human approvals, and rapid rollbacks.

    AI & LLMs· Resource

  • X post:ClaudeDevs

    How the Claude Code Team Redesigned Code Review for AI

    Around 90 percent of Claude Code is written by Claude itself. The Anthropic engineering team shares how shifting to automated, multi-layered reviews more than doubled merged pull requests while cutting reverts by over 40 percent.

    10kbookmarks1.6mviews

    AI & LLMs· Opinion

  • Article:Ryan Peterman

    How Anthropic Builds and How Engineering Will Change Soon

    In this episode of The Peterman Pod, host Ryan Peterman interviews Thariq Shihipar from Anthropic's Claude Code team about how Anthropic uses models for internal engineering and how software engineering will evolve.

    AI & LLMs· Discussion

  • Article:Anthropic

    Anthropic Introduces Claude Fable 5 and Claude Mythos 5

    Anthropic has released Claude Fable 5 and Claude Mythos 5, two Mythos-class AI models. Fable 5 is tailored for general use with conservative safety fallbacks, whereas Mythos 5 provides unrestricted access to advanced cybersecurity capabilities for vetted defenders via Project Glasswing.

    AI & LLMs· Announcement

  • Article:Anthropic

    Anthropic Unveils Claude Opus 4.6 with 1M Token Context Window

    Anthropic has released Claude Opus 4.6, delivering major upgrades to agentic coding, multi-step planning, and long-context retrieval with a 1M token beta window, alongside new effort controls at existing API pricing.

    AI & LLMs· Announcement

  • Video

    Video:CURT

    How Anthropic Builds Software: Thariq Shihipar on Claude Code and the Future of Engineering

    In an interview with Ryan Peterman, Thariq Shihipar, an engineer on Anthropic's Claude Code team, discusses how engineering practices are evolving alongside advanced AI models. At Anthropic, developers treat Claude as a thought partner operating at higher levels of abstraction, with manual IDE coding largely diminishing. Shihipar discusses the critical role of harness engineering, loop engineering workflows, verification strategies to prevent incidents, and why foundational technical knowledge remains vital even as coding becomes increasingly automated.

    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.