5-MINUTE WOW · OPENAI AGENTS SDK

OpenAI Agents SDK + NocturnusAI

The problem: Every tool call re-sends the prior conversation into the agent's context.

Without NocturnusAI

1,259

tokens / turn

→

With NocturnusAI

221

tokens / turn

Reduction

5.7×

82% fewer tokens

Measured on our 15-turn product-support benchmark against Claude Opus 4. The compression happens at the Nocturnus layer — it's framework-agnostic. Run bench.py against your own workload.

Copy-paste install

pip install nocturnusai openai-agents
docker run -d -p 9300:9300 ghcr.io/auctalis/nocturnusai:latest
export OPENAI_API_KEY=sk-...

2-minute demo

CRM-lookup agent resolving an SLA-credit ticket.

What changes in your code

get_nocturnusai_tools() returns a list of Agents-SDK-compatible tools. Pair them with process_turns() before each agent run and the agent sees only what changed.

from nocturnusai import SyncNocturnusAIClient
from nocturnusai.openai_agents import get_nocturnusai_tools
from agents import Agent, Runner

with SyncNocturnusAIClient("http://localhost:9300") as client:
    # Compress the ticket thread once
    client.process_turns(
        turns=[
            "User: Globex asking for SLA credits, claims 14h downtime.",
            "Tool sla_calculator: breach=true, credit_owed=$77500.",
            "Tool contract_check: auto_approval_limit=$100k.",
        ],
        scope="sla-ticket-8812",
        session_id="sla-ticket-8812",
    )

    # Hand the knowledge-query tools to the agent
    agent = Agent(
        name="sla-resolver",
        instructions="Resolve SLA credit tickets. Query the KB (scope='sla-ticket-8812').",
        tools=get_nocturnusai_tools(client),
    )

    result = Runner.run_sync(agent, "Is this ticket eligible for auto-approval?")
    print(result.final_output)
    # → "Yes — $77,500 is under the $100,000 auto-approval threshold."

Why it works

The agent never sees the raw ticket thread. It sees facts like credit_owed(globex, 77500) and auto_approval_limit(100000) — and backward-chains from eligible_for_auto_approval(globex) to the answer. The context stays flat turn-over-turn.

What’s in the repo

knowledge_agent.py — full agent wiring against a 5-turn SLA ticket
bench.py — run a 15-turn conversation with and without NocturnusAI against your own OpenAI key
requirements.txt

Run it yourself

GITHUB

Full example directory

examples/openai-agents/

METHODOLOGY

Benchmark details

How the numbers were measured, and how to reproduce them.

NEXT EXAMPLE

CrewAI

Each crew member receives all prior crew outputs as context — token cost scales with crew size.