In short
AI agents carry out tasks in many steps, calling models and tools over and over, so a single request can use many times the tokens of a chat. Multi-agent systems multiply this, and when agents loop or talk past each other the tokens are wasted. Budgets, step limits, loop detection and right-sized models keep agents useful without burning energy for nothing.
Key points
- An agent makes many model calls per task, so its token use, and energy use, can be many times that of a single chat.
- Multi-agent systems multiply token use again, which pays off only for tasks valuable and complex enough to need it.
- Agents can get stuck in loops or repeat work, consuming tokens without making progress.
- Hard limits on steps, tokens, time and spend, plus monitoring that can stop a stuck run, are basic guardrails.
- Right-sizing means using a single agent, a small model or no agent at all when the task allows.
AI agents are systems that work towards a goal in several steps, choosing which tools to call and what to do next. They can book, search, write code, test it and try again. That autonomy is useful, but each step is another call to a model, and the costs add up in ways a single chat does not.
What makes agents expensive?
An agent calls a model many times per task, and each call usually includes the growing history of the task so far. Token use, and so energy use, grows with every step, tool call and retry.
A token is a chunk of text, roughly part of a word, that a model reads or writes. Every token processed costs computation. In a chat, one question produces one answer. In an agent, one request may trigger dozens of model calls: planning, searching, reading results, writing, checking, and revising.
Anthropic described this in a 2025 account of building its multi-agent research system. It reported that agents typically used about four times more tokens than chat interactions, and multi-agent systems about fifteen times more. It also found that, on its browsing benchmark, token usage by itself explained 80% of the variance in performance: the systems did better largely because they spent more.
That is a fair trade for a hard, valuable task. It is waste for a simple one.
Is there evidence on agents' energy use?
Early research points the same way. Kim, Shin, Chung and Rhu (2025) compared agentic, multi-turn workflows with single-turn inference from the point of view of data-centre infrastructure. They found that agents used substantially more energy, that the gains in accuracy showed diminishing returns as more computation was spent, and that the infrastructure costs of scaling agents this way could become unsustainable.
Measured figures for real deployments are still scarce, because providers rarely publish them. See measurement is the bottleneck.
How do agents waste tokens?
By doing work that does not move the task forward. Common patterns include:
- Loops. An agent repeats the same action, or two agents keep handing a problem back and forth.
- Over-checking. Every step is reviewed by another model call, even where a simple test would do.
- Context bloat. Whole documents, logs or tool outputs are carried through every later step.
- Retry storms. A failure triggers repeated retries, sometimes on ever larger models.
- Over-delegation. A task is split across many agents when one would have done.
Cemri and colleagues (2025) analysed more than 1,600 annotated traces from seven multi-agent frameworks and identified 14 distinct ways these systems fail, grouped into three categories: problems with how the system and task are specified, misalignment between agents, and failures to verify the task. Each failure mode is also a way of spending tokens without getting a result.
A 2026 UK green AI sandpit of mathematicians, computer scientists and industry discussed the same problem: swarms of agents can trap themselves, sending each other meaningless output and burning through hundreds of thousands of tokens. The open research question it raised is how to detect, mathematically and early, that a group of agents is stuck, so that someone or something can stop it.
Does this matter commercially as well as environmentally?
Yes; cost and energy rise together. Gartner predicted in June 2025 that over 40% of agentic AI projects will be cancelled by the end of 2027, due to escalating costs, unclear business value or inadequate risk controls. Token waste is a direct contributor to the first of those.
What guardrails keep agents in check?
Hard limits, monitoring and a way to stop. These are basic engineering controls, and most agent frameworks support some of them.
| Guardrail | What it does |
|---|---|
| Step limit | Caps the number of model calls in a run. The OpenAI Agents SDK, for example, has a max_turns setting that stops a run when it is exceeded. |
| Token or spend budget | Stops or escalates a run when it passes a set cost. |
| Time limit | Ends runs that take far longer than expected. |
| Loop detection | Flags repeated actions or near-identical messages between agents. |
| External observer | A separate, cheap process or a person that watches progress and can stop the run: the "red button". |
| Context trimming | Summarises or drops old history so each call reads fewer tokens. |
Limits should fail safely: when a budget runs out, the agent should stop and report what it has, not silently retry.
What does right-sizing an agent mean?
Using the least elaborate system that does the job. Before building a multi-agent system, work down this list:
- No agent. Can a fixed script, a search or a single model call do it?
- One call with tools. A single, well-prompted model call that can use a tool is often enough.
- One agent. A single agent with a step limit.
- Multiple agents, only for tasks that are broad, valuable and genuinely parallel.
Within any of these, use small models for simple sub-tasks such as classification, extraction and formatting, and keep large models for the hard reasoning. There is a catch: a model that is too small may fail and force a rerun on a larger one, and in a long chain one failure can derail later steps. Setting the threshold is a design decision worth testing. See choosing the right model and orchestration over generation.
What to do next
- Log tokens per completed task, including retries and failed runs, for every agent you run.
- Set step, token, time and spend limits on every agent before it goes into production.
- Add loop detection and a way for a person to stop a run.
- Start with the simplest design that works, and add agents only when the task needs them.
- Route simple sub-tasks to small models. See use AI efficiently.
- If you buy agentic tools, ask suppliers how they limit and report token use: see sustainable AI procurement.
Frequently asked questions
What is agentic AI?
AI systems that pursue a goal over several steps, deciding for themselves which tools to call and what to do next, rather than answering a single prompt. They often call a language model many times per task.
Why do AI agents use so many tokens?
Each step sends the growing history of the task back to the model, and agents may call tools, check their own work and retry. Multi-agent systems also pass messages between agents, each of which costs tokens.
How can I stop an AI agent wasting tokens?
Set limits on the number of steps, tokens, time and spend; detect loops and repeated actions; use small models for simple sub-tasks; and keep a person able to stop a run.
Sources
- How we built our multi-agent research system, Anthropic (2025)
- Why do multi-agent LLM systems fail?, Cemri et al. (2025)
- The cost of dynamic reasoning: demystifying AI agents and test-time scaling from an AI infrastructure perspective, Kim, Shin, Chung and Rhu, HPCA 2026 (2025)
- Gartner predicts over 40% of agentic AI projects will be canceled by end of 2027, Gartner (2025)
- Running agents (max_turns), OpenAI Agents SDK documentation (2025)
Related
Orchestration over generation
Why reusing vetted components and routing requests to the smallest capable model can cut AI's energy use, and what the evidence says so far.
Use AI efficiently: everyday habits
Everyday habits that cut the energy of using AI: clear prompts, fewer regeneration loops, care with images and video, caching and turning off defaults.
Choose the right model for the job
How to right-size AI: when to use small or specialised models, when a frontier model earns its cost, how routing works, and when not to use AI.
Measurement is the bottleneck for green AI
Why AI energy figures are hard to trust or compare: self-reported numbers, estimates not meters, different boundaries, and what good looks like.