Architectes du SI Agentique d'Entreprise

1from google.adk.agents import LlmAgent, SequentialAgent
2from google.adk.tools import agent_tool
3from tools.mcp import salesforce, confluence, jira
4 
5# --- Sub-agents: each bounded to its own MCP tools ---
6 
7classifier = LlmAgent(
8    name="DocumentClassifier",
9    model="gemini-2.5-pro",
10    instruction="Classify document by type, urgency, regulatory scope.",
11    tools=[confluence],
12    output_key="classification",
13)
14 
15extractor = LlmAgent(
16    name="DataExtractor",
17    model="gemini-2.5-pro",
18    instruction="Extract structured fields from {classification}.",
19    tools=[salesforce, jira],
20    output_key="extracted_data",
21)
22 
23validator = LlmAgent(
24    name="ComplianceValidator",
25    model="gemini-2.5-pro",
26    instruction="Validate {extracted_data} against regulatory rules.",
27    tools=[confluence],
28    output_key="validation_report",
29)
30 
31human_review = LlmAgent(
32    name="HumanGate",
33    model="gemini-2.5-pro",
34    instruction="""If validation_report.confidence < 0.85:
35      escalate to human reviewer with full context.
36      Otherwise: auto-approve and route to next step.""",
37    output_key="final_decision",
38)
39 
40# --- Orchestrator: chains sub-agents in sequence ---
41# Each agent receives the output_key of the previous one
42# The pipeline is fully observable via LangSmith traces
43 
44root_agent = SequentialAgent(
45    name="DocumentProcessingPipeline",
46    description="Enterprise intake → classify → extract → validate → gate.",
47    sub_agents=[classifier, extractor, validator, human_review],
48)