ModelRefs / Agentic Loop — AI Glossary
Agentic Loop — AI Glossary
The observe → think → act → observe cycle an agent repeats until its goal is achieved or a stopping condition is met. Also called agent loop or ReAct loop.
Overview
The agentic loop is the core execution pattern for all LLM agents. Each iteration: observe environment state, reason about next action, execute a tool call or produce output, then re-observe. The loop terminates when the agent decides the task is complete or hits a max-steps limit.
Reference details
| Topic | agents |
|---|---|
| Also known as | agent loop, ReAct loop |
| Last reviewed | 2026-06-24 |
Related terms
Example: Cost grows with the square of the step count
Each iteration resends the whole transcript so far. Start at 2,000 tokens and add 500 per step: step one sends 2,000, step fifteen sends 2,000 + 14 × 500 = 9,000. Summed across fifteen steps that is 15 × 2,000 + 500 × (0+1+…+14) = 30,000 + 52,500 = 82,500 input tokens for one task. Doubling the step budget to thirty does not double the cost — it takes the total to 277,500, more than triple, and the gap widens the longer the loop runs. This is why a step cap is a cost control as much as a safety net, and why compacting or summarising old steps is what actually makes long loops affordable.
Commonly confused with
The loop is the execution pattern; ReAct is one prompting scheme for the reasoning step inside it; an agent is the whole system, including tools, memory and permissions. Frameworks bundle all three, which makes it easy to blame “the agent” for what is a loop-control problem — no stopping condition, no step budget, no progress check.
When to use it
Reach for it when:
- Tasks whose step count genuinely depends on intermediate results
- Whenever a step budget, a stopping condition and a progress check are defined together
- With tracing enabled — loops are unreadable after the fact without per-step spans
Reach for something else when:
- Fixed pipelines, where a chain is cheaper, faster and testable
- Without a cap: a loop that chooses its own next step can choose badly many times
- Where every iteration is irreversible and unattended, with no human interrupt point
Continue your research
Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Agentic Loop — AI Glossary.
Frequently asked questions
What is Agentic Loop?
The observe → think → act → observe cycle an agent repeats until its goal is achieved or a stopping condition is met.
Is Agentic Loop the same as agent loop?
Yes — agent loop, ReAct loop are common aliases for Agentic Loop.
What concepts are related to Agentic Loop?
Closely related concepts include agent, react agent, tool use.