In June 2026, a new concept entered the vocabulary of everyone building with AI coding agents: loop engineering. The term, popularized by Addy Osmani, describes a shift from manually prompting agents to designing the systems that prompt them — autonomous loops that generate, evaluate, and iterate on work until predefined goals are met. As firms such as 650 Group have been tracking in their AI software development research, the speed at which this concept moved from viral post to established practice is unusual even by AI industry standards.
The trigger was a series of developments concentrated in a two-week window. Andrej Karpathy joined Anthropic in late May and, shortly after, a CLAUDE.md file attributed to him began circulating that contained four behavioral rules for Claude Code: Think Before Coding, Simplicity First, Surgical Changes, and Goal-Driven Execution. The file was posted on GitHub and within days had been forked, adapted, and embedded into thousands of development workflows. A community gist merged the CLAUDE.md rules with a new LOOPS.md framework that formalized a gather-reason-act-verify-repeat cycle for autonomous agent work.
At roughly the same time, Boris Cherny, the creator and head of Claude Code at Anthropic, revealed in interviews with Fortune and Lenny’s Newsletter that he manages a few hundred agents some mornings and thousands on other days. He reported that 100% of his own code — and much of Anthropic’s — is now produced by Claude Code with heavy agent use. Major customers including Salesforce showed similar production shifts. The Claude Code system had grown to account for 4% of all public GitHub commits, as tracked by SemiAnalysis.
Osmani’s article on loop engineering, published in late June on his blog and reposted by O’Reilly, defined five building blocks: automations for scheduled discovery and triage, isolated worktrees for parallel agent execution, documented skills for project context, plugins and connectors for tool integration, and sub-agents for both ideation and verification. The framework also called for external memory to maintain persistent state across runs — a capability that Anthropic’s Claude Code would formalize as “skills” in subsequent weeks.
Anthropic’s own Claude Code team published an official guide on June 30 that defined four types of agentic loops with increasing autonomy. Turn-based loops rely on a user prompt with the agent judging completion. Goal-based loops use a /goal command with a separate evaluator model checking verifiable criteria after each turn. Time-based loops use /loop for local session intervals and /schedule for cloud routines independent of an open session. Proactive loops combine event triggers, goals, skills, and dynamic workflows for truly unattended recurring work.
The /goal and /loop commands were already shipping in Claude Code by early July. The /schedule functionality, which enables cloud-hosted routines that persist across sessions, was documented alongside the proactive loop pattern. Documentation was published at code.claude.com/docs/en/goal and code.claude.com/docs/en/routines.
The significance of this shift extends beyond coding. Loop engineering represents a change in how people interact with AI systems: from crafting individual prompts to designing architectures of autonomous agent behavior. The five building blocks Osmani described — automations, worktrees, skills, connectors, and sub-agents — are not Claude-specific. They are architectural primitives that apply across agent frameworks, including Codex, Claude Code, and Hermes Agent.
A community response was almost immediate. HK students published an 18-page PDF claiming to improve Karpathy’s loop design by 5x, accumulating significant engagement within days. Open-source repositories for loop-engineering frameworks appeared on GitHub. Anthropic published additional resources on dynamic workflows for long-running Claude for scientific computing. The CIA Red Cell playbook for Claude — prompts for adversarial analysis — was also shared publicly.
The shift from prompt engineering to loop engineering is still in its early days, but the infrastructure is assembling quickly. Anthropic has provided the primitives. Osmani has provided the vocabulary. Karpathy has provided the behavioral constraints. The question now is whether organizations will adopt these patterns as quickly as individual developers have.