Why AI agents burn your budget and how to stop it

An agent is a loop of requests

A person asks and reads. An agent loops: think, call a tool, read the result, think again. On each turn it sends the model the whole history: instructions, files read, command output.

So agent cost grows faster than the number of steps: every next step is heavier than the previous one because the history is longer.

How to cut usage if you pay per token

• Cap steps and time per task
• Compress history: keep step summaries instead of full command output
• Do not let the agent read whole files when it needs one function
• Use a fast model for simple steps and a large one for hard steps
• Cache repeated instructions if your provider supports it

Or remove the meter

Everything above fights the billing model. If an agent is your main tool, change the billing model: with a stream subscription the number of tokens does not affect the price.

At Tenzo Cloud an agent can take any number of steps; only concurrent requests are limited. For an agent with helpers we recommend 4 streams.

An agent on the OpenAI SDK: only the address changes

from openai import OpenAI

client = OpenAI(base_url="https://tenzo-cloud.com/v1", api_key="tc-your-key")

stream = client.chat.completions.create(
    model="glm-5.3",
    messages=[{"role": "user", "content": "Hello!"}],
    stream=True,
)
for chunk in stream:
    print(chunk.choices[0].delta.content or "", end="")

Helpers and parallelism

Modern agents delegate subtasks to helpers: one searches, another writes tests, a third reviews code. Each helper is another stream of requests with its own context, so token usage multiplies.

In the Tenzo Cloud chat, helpers work within your streams: the assistant sees how many are free and delegates only when there is room. Parallel work speeds things up without growing the bill.

Signs an agent is wasting work

• It reads the same file several times
• The step history has the same command failing again and again
• A 10-minute human task takes the agent an hour
• It does not stop until interrupted