route_agent.py

import asyncio
import json

from strands import Agent, tool
from strands_tools import use_llm
from strands_tools import calculator
from strands.models import BedrockModel

bedrock_model = BedrockModel(
    model_id="us.anthropic.claude-3-7-sonnet-20250219-v1:0",
    region_name='us-east-1',
    temperature=0.3,
)
SYSTEM_PROMPT = '''
You are a routing agent. Your job is to read the incoming message, determine the most appropriate destination from the allowed labels, and hand off with a minimal, useful payload.
Success: choose the correct label, include a short rationale, and provide the essentials the destination agent needs to act.
Principles:

Be decisive when confidence ≥ threshold; otherwise ask one targeted clarification or return other.

Never invent capabilities or promises; prefer safe deferral to misrouting.

Preserve privacy: don’t request secrets/PII; redact sensitive data in the payload.

Keep context tight: summarize long threads into 3–5 bullets before routing.
Output (JSON): { "label": "<one of [pricing, tech_support, partnerships, other]>", "confidence": 0.0–1.0, "rationale": "one sentence", "handoff_payload": { ...essentials } }
Threshold: default 0.7 unless configured otherwise.
'''


@tool
def clasify_intent(input: str):
    '''
    You classify an incoming message into one of the known labels.
    Success: return the best label + confidence + one-line rationale.
    Principles: be decisive when confidence ≥ threshold; if < threshold, defer to Clarify.
    '''
    CLASSIFY_INTENT_SYSTEM_PROMPT = '''
    You classify an incoming message into one of the known labels.
    Success: return the best label + confidence + one-line rationale.
    Principles: be decisive when confidence ≥ threshold; if < threshold, defer to Clarify.
    '''
    classify_agent = Agent(
        model=bedrock_model,
        tools=[use_llm],
        callback_handler=None
    )

    response = classify_agent.tool.use_llm(
        system_prompt=CLASSIFY_INTENT_SYSTEM_PROMPT,
        prompt=input,
    )
    print(response)

    return response


@tool
def route_to_pricing(input: str):

    ROUTE_TO_PRICING_AGENT_SYSTEM_PROMPT = '''
   You draft accurate pricing/plan responses using documented offerings only.
    Success: clear answer + next step (e.g., link to plans, request usage details).
    Principles: no discounts/commitments; flag unknowns instead of guessing.
    '''
    route_to_pricing_agent = Agent(
        model=bedrock_model,
        system_prompt=ROUTE_TO_PRICING_AGENT_SYSTEM_PROMPT,
        callback_handler=None
    )

    response = route_to_pricing_agent(input)
    print(response)

    return response

@tool
def route_to_tech_support(input: str):

    ROUTE_TO_TECH_SUPPORT_AGENT_SYSTEM_PROMPT = '''
   You produce helpful troubleshooting replies and safe next steps.
    Success: reproducible steps, minimal data requests, and a quick diagnostic path.
    Principles: never ask for secrets/PII; prefer redacted logs and environment basics.
    '''
    route_to_tech_support_agent = Agent(
        model=bedrock_model,
        system_prompt=ROUTE_TO_TECH_SUPPORT_AGENT_SYSTEM_PROMPT,
        callback_handler=None
    )

    response = route_to_tech_support_agent(input)
    print(response)

    return response

@tool
def route_to_partnership(input: str):

    ROUTE_TO_PARTNERSHIP_AGENT_SYSTEM_PROMPT = '''
   You handle BD/partnership inquiries and propose times to meet.
    Success: warm, concise reply with 2–3 scheduling options and a clear CTA.
    Principles: no legal/financial commitments; gather just enough context to proceed.
    '''
    route_to_partnership_agent = Agent(
        model=bedrock_model,
        system_prompt=ROUTE_TO_PARTNERSHIP_AGENT_SYSTEM_PROMPT,
        callback_handler=None
    )

    response = route_to_partnership_agent(input)
    print(response)

    return response


# Initialize our agent without a callback handler
router_agent = Agent(
    model=bedrock_model,
    system_prompt=SYSTEM_PROMPT,
    tools=[clasify_intent, route_to_pricing,route_to_partnership,route_to_tech_support],
    callback_handler=None
)

# Async function that iterates over streamed agent events


query = '''
We’d like to co-host a webinar with Educloud Academy next month.”
→ partnerships | {topic:"webinar", timeframe:"next_month"}'''


agent_response = router_agent(query)
print(agent_response)


#sample prompts

'''
Pricing

“What does the Pro plan cost for 25 seats and monthly billing?”
→ label: pricing | payload: {seats:25, billing:"monthly"}

“Is there a discount for EDU/nonprofits?”
→ pricing | {segment:"education|nonprofit"}

“Compare Starter vs Pro for API rate limits.”
→ pricing | {compare:["Starter","Pro"], concern:"rate_limits"}

“Can we pay via invoice instead of card?”
→ pricing | {payment_method:"invoice"}

“How much is Pro in XAF?”
→ pricing | {currency:"XAF"}

Tech Support

“Deploy failing: ValidationException: engine not supported in region eu-west-3.”
→ tech_support | {error:"ValidationException", region:"eu-west-3"}

“Webhook 500 when sending to /events; here’s the payload (redacted).”
→ tech_support | {endpoint:"/events", status:500}

“Your Python SDK v3.2 throws NameError after upgrade.”
→ tech_support | {sdk:"python", version:"3.2"}

“Emails from SES aren’t reaching Gmail; SPF/DKIM look fine.”
→ tech_support | {feature:"email", provider:"SES", target:"Gmail"}

“App crashes on iOS 17.5 only.”
→ tech_support | {platform:"iOS", version:"17.5"}

Partnerships

“We’d like to co-host a webinar with Educloud Academy next month.”
→ partnerships | {topic:"webinar", timeframe:"next_month"}

“Exploring a reseller agreement in West Africa—who should we talk to?”
→ partnerships | {deal_type:"reseller", region:"West Africa"}

“Can we integrate your API into our LMS and do a joint case study?”
→ partnerships | {integration:"LMS", marketing:"case_study"}

“Interested in sponsoring your AI for Kids program.”
→ partnerships | {program:"AI for Kids"}

Ambiguous → Clarify

“How do I get started?” (could be pricing, onboarding, or support)
→ confidence < 0.7 → ask: “Do you need pricing, setup help, or a partnership intro?”

“I need help with the plan for my team.” (plan = pricing? setup?)
→ clarify with 3 choices: pricing / technical setup / something else.

“We want to talk.”
→ clarify: “Is this about pricing, technical support, or partnerships?”

Other / Non-routable or Safety

“Can you share your customers’ private emails?” (policy disallowed)
→ other | rationale: policy; payload empty.

“Here’s my password: P@ssw0rd123—please fix my account.”
→ other | redact & respond with safe guidance (no secrets).

Non-English: “¿Ofrecen planes con factura anual?”
→ pricing | {language:"es", billing:"annual"}

Forward/noisy thread: “Fwd: Re: Re: pricing?? (see attached, long thread)”
→ pricing | summarize top 3 bullets before handoff.

Multi-intent: “Site down + need a quote for 50 seats.”
→ split or prioritize:

ticket A → tech_support {incident:"site_down"}

ticket B → pricing {seats:50}

If you want, I can turn these into unit tests with expected router JSON output (label, confidence, rationale, handoff_payload).
'''