Loading the page.
Documentation
Journal
Frontend
Backend
Admin
CI/CD
Loading the page.
Card · concepts · updated 13.08.2026
concept · a method of working with AI agents
A loop (agent loop) is when an AI agent is given not a request, but a process: the agent makes a pass, an external check decides "done or again", and a stop condition closes the loop. A prompt buys one turn; a loop buys a verified result. In summer 2026, this became a product: /goal, /loop, and routines in Claude Code, Automations in Codex.
4
Types
/goal /loop
Teams
summer 2026
Wave
AutoGPT 2023
Precursor
task + "done" condition
│
▼
┌── agent makes a pass
│ │
│ ▼
│ result check
│ │ │
└── again done stop
# the loop runs until a stop condition is met: a criterion, a turn limit, or cancellation▌What it gives you
What it is not
What we call it
We call it: loops. Also known as: loops, agent loop, agent cycle. Didn't stick: agent loop.
Verdict vibecoding.tech · 2026-08-13 · full glossary
Four nodes, no magic: a loop is scaffolding around an ordinary agent. The loop's strength and failure are decided in the third node.
Families · Claude Code team's taxonomy (30.06.2026): by trigger and stop condition.
| Type | What it's for | Today |
|---|---|---|
| Turn-based | standard pass: prompt, work, response | every request you make to an agent; enhanced with verification skills |
| Goal-based | a loop until a verifiable criterion | /goal: "Lighthouse ≥ 90, stop after 5 attempts" |
| Time-based | repeat on a schedule | /loop and /schedule: babysit PRs, morning summary |
| Proactive | an event wakes the agent, no human involved | incoming bug handling: triage, fix, review |
| Self-improvement | the loop improves the agent itself using failure traces | Replit setup: failure cluster, hypothesis, draft PR (07.2026) |
The idea is less than four years old, and the first attempt failed. The timeline explains what changed between 2023 and 2026: not the loop mechanic, but the judge inside it.
2023
AutoGPT and BabyAGI
the first viral loops on top of GPT-4: 100k+ stars in months and almost no shipped products - the agent checked itself and looped forever
2025.07
Ralph loop
Geoffrey Huntley: an endless bash loop feeds the agent PROMPT.md with clean context; crude, but according to the author, an MVP for a $50k contract was delivered for $297 in API costs
2025.09
The definition settled
Simon Willison: "an agent is an LLM running tools in a loop toward a goal"; Anthropic agrees with this formula
2026.05
Loops become products
Claude Code releases /loop, /goal, routines, and dynamic workflows; Codex has Automations, Cursor has background agents
2026.06
The "loop engineering" wave
Steinberger's post with millions of views and Osmani's essay give the practice a name; the creator of Claude Code: "my job is to write loops"
2026.07
Loops that learn
Replit describes a self-improvement loop: the agent reads production failures and proposes changes to itself; launch authority belongs to the engineer
Choosing a loop means choosing a trigger and a stop condition. Honestly: most tasks do not need a loop - start with a standard pass and build up.
Teams
/goal - loop until a criterion (v2.1.139+); /loop - repeat at an interval (v2.1.72+); /schedule - cloud routines
/goal: after each turn, a cheap evaluator model checks the condition against the transcript. /loop: session-scoped, auto-expires after 7 days. Routines: minimum interval 1 hour, the machine can be off.
docs: /goal · 2026-07-07
The cost
$1.3M in 30 days: the loop token-burning record - 603 billion tokens, ~100 agents (OpenClaw, 2026)
Steinberger, creator of OpenClaw: agents review PRs, find bugs, deduplicate issues; OpenAI pays the bill. An edge case, not the norm: a home loop on a subscription costs orders of magnitude less.
Tom's Hardware · 2026-07-07
From Russia
A loop is a mechanism inside an agent, not a separate service: it works wherever the agent itself works
Claude Code and Codex from Russia require foreign subscription payment; access details and payment methods are on the agent cards.
Claude Code card · 2026-07-07
There is maximum hype around loops in both directions. A breakdown of the most persistent claims:
"A loop is an agent without people"
a person remains at four points: choosing hypotheses, architecture, verification criteria, and merge authority. Replit explicitly keeps launch approval with the engineer (07.2026)
"This already existed - AutoGPT"
AutoGPT judged itself and got stuck in loops; 2026 loops rely on external verification - tests, CI, an evaluator model. The difference is like between "I checked myself" and "someone else administered the exam"
"Loops burn money for nothing"
they do, if there are no stop conditions: the record is $1.3M/month (OpenClaw). Managed with three controls: action cap, interval based on change velocity, and a cheap model for verification
"Started it and left"
loop quality = verification quality. If you cannot write down how to verify the result, do not start a loop: you will get "looks done" on repeat
"This is a Claude Code feature"
the mechanism is universal: Codex Automations, Cursor background agents, Copilot coding agent - every major agent has its own form of loop (2025-2026)
Words from adjacent worlds get confused constantly. The order is not accidental: each next level wraps the previous one.
| Term | What it is | Example |
|---|---|---|
| ·Prompt | one request: a task in one pass | "fix the button and show the diff" |
| ⊃Agent | an LLM running tools in a loop toward a goal | Claude Code, Codex, Cursor |
| ⊃⊃Loop | a loop around an agent: trigger + verification + stop condition | /goal, /loop, Ralph loop in bash |
| ⊃⊃⊃Workflow | orchestration of many agents by a script | dynamic workflows: 20 reviewers in parallel |
| ⊃⊃⊃⊃Routine | a loop that lives in the cloud on a schedule | /schedule: bug summary every morning |
It is one concept. On the site - cycles (verdict 2026-08-13): a living Russian word, and that is how people search too - “agent loop” persists in Yandex and Google suggestions (measurement 07.2026). Practitioners say “loops” (a transliteration of loops); “agent loop” did not stick.
A prompt buys one run: the agent does the work and stops, and you check it. A loop buys a process: the agent runs again and again, verification is built in, and it stops by condition, not by the model’s mood.
No. A human remains at four points: which tasks deserve a loop, how to fix things architecturally, what counts as verification, and what to ship. Everything between those points - research, edits, runs - the machine does itself.
From almost zero to $1.3M per month (the OpenClaw record, ~100 agents). Cost control: a turn cap (“stop after 20”), an interval based on the rate of change in what you are monitoring, and a cheap model to check the condition.
The mechanics are similar, but the judge is different. AutoGPT self-checked and got stuck in loops; 2026 loops rely on external validation: tests, CI, screenshots, a separate evaluator model. Plus, the models themselves can now deliver multi-hour tasks.
Yes: a loop is a mechanism inside an agent, not a separate service. From Russia, the question comes down to access to the agent itself: Claude Code and Codex require foreign subscription payment. Access and payment details are on the agent cards.
Take a task with a verifiable “done” state and give it to /goal with an attempt cap: “raise the homepage Lighthouse to 90+, stop after 5 attempts.” A verifiable criterion instead of “make it good” is 80% of success.
The old-school version: an infinite bash loop that feeds the agent the same PROMPT.md with a clean context on every iteration. Crude, but it worked before built-in /loop and /goal - and explains the essence of a loop better than any theory.
Related concepts
Vibecodingthe practice that loops build on: first learn one pass
MCPthe tool outlet: a loop that sees only files - a small loop
Claude Codean agent where loops became a product: /goal, /loop, routines
More on the topic
Entry · loops · facts · 10 (source and date for each) · card edited 2026-08-13 · facts verified 2026-07-07
Where to go next
New cards, price and limit changes, ranking shifts - by email, no spam.