atomic-agents | Doubleword Inference API

atomic-agents uses Instructor (which wraps the OpenAI client), so it supports custom endpoints by passing a configured client.

Install

pip install atomic-agents openai instructor

Configure

import instructor
from openai import OpenAI
from atomic_agents import AtomicAgent, AgentConfig

client = instructor.from_openai(
    OpenAI(
        base_url="https://api.doubleword.ai/v1",
        api_key="{{apiKey}}",
    )
)

agent = AtomicAgent(
    config=AgentConfig(
        client=client,
        model="{{selectedModel.id}}",
    )
)

response = agent.run(agent.input_schema(chat_message="Say hello."))
print(response.chat_message)

The base_url and api_key are standard openai.OpenAI constructor parameters. instructor.from_openai() wraps the client for structured output support.

Prompt caching

AtomicAgent builds its system prompt as a string. One subclass turns it into a marked block:

import instructor
from openai import OpenAI
from atomic_agents import AtomicAgent, AgentConfig

client = instructor.from_openai(
    OpenAI(
        base_url="https://api.doubleword.ai/v1",
        api_key="{{apiKey}}",
    )
)


class CachedAgent(AtomicAgent):
    def _build_system_messages(self):
        messages = super()._build_system_messages()
        if messages:
            messages[0]["content"] = [{
                "type": "text",
                "text": messages[0]["content"],
                "cache_control": {"type": "ephemeral", "ttl": "1h"},
            }]
        return messages


agent = CachedAgent(
    config=AgentConfig(
        client=client,
        model="{{selectedModel.id}}",
        system_prompt_generator=large_system_prompt,
    )
)

agent.register_hook("completion:response", lambda r: print(r.usage))
response = agent.run(agent.input_schema(chat_message="What is 2 + 2?"))

large_system_prompt is your SystemPromptGenerator and needs to clear the ~1024-token floor. Keep the default Mode.TOOLS because instructor's JSON modes rewrite the first content block. Read cache counts in the completion:response hook since instructor drops them from response.usage. See the prompt caching guide.