AI agents

One user action becomes dozens of model calls. That changes which parts of the platform matter.

What is different about agents

• The cost is in the volume of calls, not any single one.
• When a run goes wrong you need to inspect one step, not a daily total.
• An unattended loop needs a ceiling it cannot exceed.

Per-step visibility

Keep the X-Astrodyne-Request-Id from each call next to your own trace id. When a run misbehaves you can open that exact step in the Request Explorer and see the model, the tokens, the latency and the charge.

import os
from openai import OpenAI

client = OpenAI(base_url="https://api.astrodyne.ai/v1",
                api_key=os.environ["ASTRODYNE_API_KEY"])

def step(messages, *, trace_id):
    raw = client.chat.completions.with_raw_response.create(
        model="YOUR_MODEL_ID",
        messages=messages,
    )
    # Store this next to your own trace so a bad step is findable later.
    log(trace_id=trace_id,
        astrodyne_request_id=raw.headers.get("X-Astrodyne-Request-Id"))
    return raw.parse()

Giving a loop a ceiling

• Run agents on a dedicated API key with its own monthly spending limit.
• The prepaid balance is the absolute ceiling — there is no credit line to overrun.
• Cap your own iteration count. The platform cannot know that step 40 is a loop.

Choosing models per step

Plan steps and final answers often warrant a frontier model; classification and routing steps rarely do. Because the model is one field in the request, mixing them costs nothing structurally. Astrodyne runs the model you name — it does not choose one for you.