duetmachines

Send your AI agent's production conversations and business outcomes to Duet Machines, which grades them against your expert's standard, builds tests from them and proves cheaper models on them.

pip install duetmachines

The client posts from a background thread and never raises into your code: if Duet Machines is unreachable, your agent keeps working and the conversation is logged as not sent.

Conversations

from duetmachines import DuetMachines

duet = DuetMachines(api_key=os.environ["DUET_API_KEY"], agent="<agent id from Traffic › Connect>")

convo = duet.conversation(
    ticket.id,                      # your id; posting again replaces the conversation
    user_id=customer.id,            # links outcomes that arrive later
    version="2026.11.1",            # your release, to compare versions
    config={"instructions": SYSTEM_PROMPT, "tools": TOOLS, "model": "gpt-5.4-mini"},
)
convo.customer(user_message)
convo.tool("lookup_order", {"order_id": "A123"}, {"status": "shipped"}, duration_seconds=0.4)
convo.agent(reply, usage={"input_tokens": 812, "output_tokens": 64, "model": "gpt-5.4-mini"})
convo.end()

usage on agent turns prices what each conversation costs, so savings can be projected. config lets graders and tests judge each conversation against the instructions and tools it ran with.

Outcomes

Send business outcomes whenever they happen, from your billing system, CRM or help desk:

duet.outcome("payment_received", id=payment.id, user_id=customer.id, value=149.00, currency="USD")
duet.outcome("ticket_reopened", id=f"reopen-{ticket.id}", conversation=ticket.id)

Declare each outcome name on the agent's Traffic › Outcomes page first.

OpenAI Agents SDK

pip install "duetmachines[openai-agents]"

from agents import add_trace_processor
from duetmachines.openai_agents import DuetMachinesProcessor

add_trace_processor(DuetMachinesProcessor(duet, version="2026.11.1"))

Every agent run is posted as a conversation with no other code.

Docs: https://platform.duetmachines.com/docs