Usage for Agents

An agent loop calls the model repeatedly — planning, acting, observing — so small inefficiencies compound fast. These patterns keep your loops predictable and cheap when running against the OpenDunes API.

Retain the client across turns#

Instantiate the SDK client once and reuse it for every call in the loop. Creating a new client per iteration adds TLS handshake overhead and makes it harder to share retry state.

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://opendunes.com/api/v1",
    api_key=os.environ["OPENDUNES_API_KEY"],
)

def agent_loop(goal: str):
    messages = [{"role": "user", "content": goal}]
    while True:
        resp = client.chat.completions.create(
            model="google/gemini-2.5-pro",
            messages=messages,
        )
        reply = resp.choices[0].message
        messages.append(reply)
        # ... parse reply, call tools, append tool results, break when done

Stream inside loops#

Streaming keeps each turn responsive even when a step produces a long answer. The final chunk includes usage so you can track token spend per turn.

with client.chat.completions.stream(
    model="meta-llama/llama-4-maverick",
    messages=messages,
) as stream:
    for chunk in stream:
        delta = chunk.choices[0].delta.content or ""
        print(delta, end="", flush=True)
    usage = stream.get_final_completion().usage

Watch your balance#

The X-Balance-Available response header carries your remaining balance in micro-DA after each call. Check it and halt the loop before you run dry — a 402 insufficient_credits mid-loop is harder to recover from than a graceful stop.

import httpx

# Use the raw HTTP response to read headers alongside the SDK response
# (or switch to the httpx client if you need header access natively)

See Response Headers for the full header reference.

Handle 429 with backoff#

The default rate limit is 60 requests per minute per key. Agent loops can saturate that quickly if steps are fast. Respect the Retry-After header rather than spinning.

import time

def call_with_backoff(client, **kwargs):
    while True:
        try:
            return client.chat.completions.create(**kwargs)
        except Exception as e:
            if hasattr(e, "status_code") and e.status_code == 429:
                retry_after = int(getattr(e, "response", {}).headers.get("Retry-After", 5))
                time.sleep(retry_after)
            else:
                raise

Keep system prompts stable#

Rewrite the system message only when the agent's role changes. Keeping it constant across turns improves model consistency and makes logs easier to read in your dashboard.

Log the X-Request-Id#

Each response includes an X-Request-Id header (UUID v7). Store it alongside the turn index so you can correlate agent traces to specific API calls when debugging.

Also see#

• Dev Tools — request logging and trace IDs.
• Streaming — full streaming reference.
• Errors — error codes and retry guidance.