Breaking

Anthropic Publishes Official Guide to Designing Loops in Claude Code

July 6, 2026 at 15:09 EDT

AI Agents
Software Dev & Coding

On July 6, 2026, Anthropic released an official guide systematizing how to design "loops" in Claude Code, its AI coding agent. It reframes the practice from tuning individual prompts ("prompt engineering") toward designing the mechanism by which an agent repeats work until a stop condition is met ("loop engineering"), laying out loop types and practical best practices.

Claude Code · Agent Design

Designing "Loops": Agents That Repeat Work Until Done

A new guide reframes agent building around loops — cycles that continue until a stop condition is met — sorted into four types, as developer attention shifts from writing single prompts to designing autonomous, long-running work.

The definition

A loop is "a cycle in which an agent repeats work until a stop condition is met" — organized into four kinds by use case.

Turn-based

Manual prompt

Short, non-recurring tasks; stops when Claude judges the task complete.

Goal-based /goal

Manual prompt

Verifiable end criterion; an evaluator model confirms completion.

Time-based /loop /schedule

Set interval

Recurring work and external monitoring at fixed intervals.

Proactive

Event / schedule

Runs with no human present — bug triage, dependency updates.

The goal-based loop cycle

Define verifiable criterion

→

Agent does work

→

Evaluator model checks

→

Repeat until met, else stop

An objective evaluator — e.g. the number of passing tests — decides when the loop ends, keeping complex tasks under control.

Keeping code quality

Maintain a clean codebase
Self-verify via SKILL.md
Review by a second agent
Use the /code-review skill

Managing token cost

Assign cheaper models to small tasks
Set clear stop criteria
Run small pilots before production
/usage breaks down consumption

The practical focus

Developers welcome the clear separation of loop types and the objective evaluator in /goal — while asking for support beyond Claude models and flagging that usage and cost management will be the real test of running autonomous, repeated work.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.