OpenAI Agents Human Approval

OpenAI Agents SDK can pause for approval. But who reminds the human? Who escalates? AXME adds the missing pieces.

Alpha - Built with AXME (AXP Intent Protocol). cloud.axme.ai - contact@axme.ai

---

The Problem

OpenAI Agents SDK lets you build powerful agents with tool use and handoffs. But when your agent processes a $2,500 refund that needs finance approval before issuing - the SDK has no built-in way to pause, notify a human, remind after 5 minutes, escalate after 30, and resume when approved.

OpenAI agent processes refund request
  -> Determines eligibility (structured, reliable, great!)
  -> Now what? Who approves this before the money moves?

Options without AXME:
  1. Poll a database in a loop? (fragile, no reminders, crashes lose state)
  2. Send a Slack message and hope? (no timeout, no escalation, no audit trail)
  3. Build webhook + queue + DB + reminder system? (weeks of infrastructure)

Open issues confirm this gap:

• openai-agents-python #378 - Human-in-the-loop approval

---

The Solution

# OpenAI Agents SDK does the reasoning
from agents import Agent, Runner

refund_agent = Agent(name="Refund Processor", instructions="Analyze refund eligibility...")
result = Runner.run_sync(refund_agent, refund_prompt)

# AXME handles the human gate
intent_id = client.send_intent({
    "intent_type": "intent.finance.refund_approval.v1",
    "to_agent": "agent://myorg/production/openai-refund-processor",
    "payload": {"refund_id": "REF-2026-3391", "eligibility": result.final_output},
})
approval = client.wait_for(intent_id)  # reminder in 5 min, timeout in 1h

OpenAI does the reasoning. AXME does the waiting.

---

Quick Start

pip install axme openai-agents
export AXME_API_KEY="your-key"   # Get one: axme login
export OPENAI_API_KEY="sk-..."   # For OpenAI Agents LLM calls

from agents import Agent, Runner
from axme import AxmeClient, AxmeClientConfig
import os

# 1. OpenAI Agents: process refund
refund_agent = Agent(
    name="Refund Processor",
    instructions="Analyze refund eligibility based on order history and policy.",
)

result = Runner.run_sync(refund_agent, "Refund REF-2026-3391: $2499.99 for service_outage")
print(f"Agent assessment: {result.final_output}")

# 2. AXME: human approval before issuing refund
client = AxmeClient(AxmeClientConfig(api_key=os.environ["AXME_API_KEY"]))
intent_id = client.send_intent({
    "intent_type": "intent.finance.refund_approval.v1",
    "to_agent": "agent://myorg/production/openai-refund-processor",
    "payload": {
        "refund_id": "REF-2026-3391",
        "amount": 2499.99,
        "assessment": result.final_output,
    },
})

# Agent suspends. Reminder in 5 min. Timeout in 1 hour.
approval = client.wait_for(intent_id)
print(f"Decision: {approval['status']}")

---

Before / After

Before: OpenAI Agents with manual polling

# OpenAI agent processes the refund...
result = Runner.run_sync(refund_agent, refund_prompt)

# ...but how does finance approve it?
print(result.final_output)        # Print to console?
input("Press Enter to approve...")  # Block the process?

# No reminders. No timeout. No escalation.
# If the approver is in a meeting, the refund is stuck forever.

After: AXME adds reminders and escalation

result = Runner.run_sync(refund_agent, refund_prompt)

intent_id = client.send_intent({
    "intent_type": "intent.finance.refund_approval.v1",
    "to_agent": "agent://myorg/production/openai-refund-processor",
    "payload": {"refund_id": "REF-2026-3391", "assessment": result.final_output},
})
approval = client.wait_for(intent_id)
# Reminder in 5 min. Escalation in 30 min. Timeout in 1h.

---

What Each Component Does

Component	Role	Framework
agent_openai.py	Process refund with OpenAI Agents SDK	OpenAI Agents
agent.py	Simplified agent (no LLM, tests AXME flow)	AXME SDK
initiator.py	Send refund intent, observe lifecycle	AXME SDK
scenario.json	Scenario definition (agent + human approval)	AXME CLI

OpenAI Agents SDK does the thinking. AXME does the infrastructure.

---

How It Works

+-----------+  send_intent()   +----------------+  process    +-----------+
|           | ---------------> |                | ----------> |           |
| Initiator |                  |   AXME Cloud   |             |  OpenAI   |
|           | <- wait_for() -- |   (platform)   | <- result   |  Agents   |
|           |                  |                |             |           |
|           |                  |  WAITING for   |             +-----------+
|           |                  |  human approval|
|           |                  |                |  approve    +-----------+
|           |                  |                | <---------- |           |
|           | <- COMPLETED --- |  - remind 5m   |             |  Finance  |
|           |                  |  - timeout 1h  |             |   Team    |
+-----------+                  +----------------+             +-----------+

---

Run the Full Example

Prerequisites

curl -fsSL https://raw.githubusercontent.com/AxmeAI/axme-cli/main/install.sh | sh
axme login
pip install -r requirements.txt

Terminal 1 - submit the scenario

axme scenarios apply scenario.json

Terminal 2 - start the agent

# macOS
cat ~/Library/Application\ Support/axme/scenario-agents.json | grep -A2 openai-refund-demo

# Linux
cat ~/.config/axme/scenario-agents.json | grep -A2 openai-refund-demo

# Simplified agent (no LLM required)
AXME_API_KEY=<agent-key> python agent.py

# Full OpenAI Agents SDK agent (requires OPENAI_API_KEY)
AXME_API_KEY=<agent-key> OPENAI_API_KEY=sk-... python agent_openai.py

Terminal 1 - approve

axme tasks approve <intent_id>

Verify

axme intents get <intent_id>
# lifecycle_status: COMPLETED

---

Related

• AXME - project overview
• OpenAI Agents SDK - agent framework
• Async Human Approval for AI Agents - framework-agnostic async approval
• AXME Examples - 20+ runnable examples

---

Built with AXME (AXP Intent Protocol).