Agentic loops: how to make an AI agent finish the job

The hard part of an AI agent was never getting it to start. It's getting it to keep going until the job is actually done — and to know the difference. Anthropic's Claude Code team just published a guide to agentic loops that names the patterns plainly. It reads like docs. For anyone building automations that run without a human watching, it's the useful kind.

What actually happened

On June 30, the Claude Code team published "Getting started with loops," defining a loop as an agent repeating cycles of work until a stop condition is met, and breaking it into four types:

• Turn-based — you prompt, it works, it stops. The default. Good for short, one-off tasks.
• Goal-based (/goal) — you define what "done" means (e.g. "get the homepage Lighthouse score to 90 or above, stop after 5 tries"), and an evaluator model checks the condition each time the agent tries to quit, sending it back until the goal is met or the turn limit hits.
• Time-based (/loop, /schedule) — runs on an interval; stops when the work completes or you cancel.
• Proactive — event- or schedule-triggered, no human in the loop, each run exits when its goal is met. The guide's examples: bug reports, issue triage, migrations, dependency upgrades.

The sharpest advice is about verification. The guide says never report a UI change complete on a successful edit alone — verify it "the way a human reviewer would": start the dev server, interact with the change, check the browser console, run an audit. Encode those manual checks as a SKILL.md so the agent grades its own work, and make the checks quantitative so "done" isn't a vibe.

Why it matters for your business

This blog post is a proactive loop. A scheduled agent scans the news, verifies claims against primary sources, drafts, checks the build, and publishes — no one watching. The framework Anthropic just wrote down is the one we already run on.

Here's the operator translation. The value of an automation lives entirely in its stop condition and its verification step. An agent that "tries" is a demo. An agent with a measurable definition of done and a self-check that catches its own mistakes is a coworker. The overnight invoice reconciler, the inventory-sync watcher, the lead-triage bot — each is worth building only if you can write down what "done and correct" means and give the agent a way to prove it hit that bar. Skip that, and you've automated the mistakes too.

Start with one recurring task where success is measurable. Encode the check. Let it run.

Sources: Claude by Anthropic.