AI Agent2026-05-17
Agent Loop
Also known as: Agent Loop / エージェントループ / 推論ループ
The cycle an AI agent repeats — reason, call a tool, receive results, re-reason — until a goal or stop condition is reached.
Overview
The loop: LLM generates a reasoning trace and tool-call JSON → host executes → result appended to context → LLM re-reasons. Repeats until done or a budget is exhausted.
Loop depth, token budgets, and error handling directly determine agent reliability. Claude Code's Hooks let developers inject logic at each iteration boundary.
Related Columns
Software Development
Complete Guide to Claude Code Hooks — Maximizing Development Efficiency Through Automation and Workflow Customization
A comprehensive guide to Claude Code Hooks covering all event types and practical applications. Learn implementation patterns for auto-formatting, lint checks, security protection, and automated testing using SessionStart, PreToolUse, and PostToolUse events.
Software Development
Claude Code Complete Guide 2026 — The New Era of AI Agent Coding
Claude Code by Anthropic is an agent-based development tool that autonomously codes by understanding entire codebases. Available via CLI, IDE, Desktop, and Web. Complete 2026 guide covering Opus 4.6, MCP, Agent Teams, and advanced features.
Software Development
AI Agents in 2026: How Businesses Are Automating Work Beyond Chatbots
In 2026, AI agents are moving from pilot projects to production deployments. Learn how they differ from chatbots and RPA, which business processes benefit most, and a practical adoption roadmap for SMBs.
AI
Loop Engineering Deep Dive — The June 2026 Successor to Prompt / Context / Harness Engineering, Crystallized by Anthropic's Boris Cherny ('I don't prompt Claude anymore — I write loops'), Named and Codified by Addy Osmani, with Six Building Blocks (Automations, Worktrees, Skills, Plugins, Maker-Checker Sub-agents, Durable State) Mapped Onto Claude Code's Existing Feature Set
A primary-source deep dive on **Loop Engineering**, the June 2026 AI-engineering trend named and codified by Google Chrome DevRel lead **Addy Osmani** in his ["Loop Engineering" blog post](https://addyosmani.com/blog/loop-engineering/) and elevated to industry attention by Anthropic Claude Code lead **Boris Cherny's** quote — **"I don it prompt Claude anymore. I have loops running that prompt Claude and figuring out what to do. My job is to write loops."** ([reported by The New Stack](https://thenewstack.io/loop-engineering/)). Covers the four-generation lineage: Prompt Engineering (2022-2024) → Context Engineering (2025, coined by Shopify CEO Tobi Lütke, formalized in [Anthropic's Effective Context Engineering for AI Agents](https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents)) → Harness Engineering (early 2026) → **Loop Engineering (June 2026 onwards)**. Grounded in Peter Steinberger's seed phrase — **"you should be designing loops that prompt your agents"** — the column maps out the six building blocks: (1) Automations / Trigger (timer- or event-driven heartbeats), (2) Worktrees (isolated git checkouts to prevent parallel sub-agent collisions), (3) Skills (SKILL.md / CLAUDE.md to externalize intent and reduce "intent debt"), (4) Plugins / Connectors via MCP (execution permissions), (5) Maker / Checker Sub-agents (separating generation from verification), and (6) Durable State (memory belongs on disk, not in context). Explains Inner Loop vs Outer Loop, how Claude Code's `/goal`, Automations, Worktrees, Skills, and Sub-agents constitute a ready-made Loop Engineering toolkit, the surge of Japanese coverage on Qiita / Zenn / DevelopersIO / note / OptiMax, and the five major risk vectors: Cognitive Surrender (Osmani's central warning), Loop Brittleness, Verifier mis-grading, HITL approval fatigue, and runaway-loop cost explosion.
Related Terms
Feel free to contact us
Contact Us