Specialized Agents Directory (23 Agents)

Overview

The platform uses a swarm of 23 specialized artificial intelligence agents built using the Pydantic AI framework (src/agents/). Each agent is configured with dependency injection (RunContext[GxpAgentDeps]), structured output validation models, and access to zero-trust hybrid retrieval tools.

Master Directory of 23 Specialized Agents

#	Agent Class	Specialization	Governing Regulations	Primary Output Schema
1	RequirementsTraceabilityAgent	Traceability Engineer	GAMP 5, 21 CFR 11.10(a)	TraceabilityMatrixResponse
2	RiskAssessmentAgent	GAMP 5 & CSA Strategist	FDA CSA, GAMP 5, ISO 14971	RiskAssessmentResponse
3	ValidationTestingAgent	Protocol Specialist	21 CFR 11.10, Annex 11 Clause 4	ValidationTestingResponse
4	DeviationRcaAgent	Quality Incident & RCA Lead	21 CFR 211.192, ISO 13485:8.5.2	DeviationRcaResponse
5	CapaChangeControlAgent	Continuous Improvement Lead	ISO 13485:8.5.2 / 8.5.3	CapaChangeControlResponse
6	DataIntegrityAuditAgent	21 CFR Part 11 Auditor	21 CFR Part 11, ALCOA+	DataIntegrityAuditResponse
7	PeriodicReviewAgent	Lifecycle Governance Lead	Annex 11 Clause 11, GAMP 5	PeriodicReviewResponse
8	AssetManagementAgent	Equipment & Metrology Sentry	ISO 13485:7.5.1, GAMP 5	AssetManagementResponse
9	CmmsMaintenanceAgent	Maintenance Governance Lead	21 CFR 211.67, ISO 13485:6.3	CmmsMaintenanceResponse
10	BatchRecordReviewAgent	Batch Release & RFT Sentry	21 CFR 211.188, Annex 16	BatchRecordReviewResponse
11	OosInvestigationAgent	QC Analytical OOS Lead	FDA OOS Guidance, ISO 17025	OosInvestigationResponse
12	GxpTrainingCompetencyAgent	Curriculum & Training Sentry	21 CFR 211.25, Annex 11 Clause 2	GxpTrainingResponse
13	ItHelpdeskAgent	GxP Service Desk Specialist	21 CFR 11.10(d), 11.300	HelpdeskTicketResponse
14	DocumentAuthoringAgent	Quality & CSV Document Drafter	ISO 13485:4.2, GAMP 5	DocumentAuthoringResponse
15	SupplierQualityAgent	Supplier Purchasing Lead	ISO 13485:7.4, FDA QMSR	SupplierQualityResponse
16	ComplaintVigilanceAgent	Post-Market Vigilance Lead	ISO 13485:8.2.2, 21 CFR 803	ComplaintVigilanceResponse
17	ApqrQualityMetricsAgent	Annual Product Review Sentry	21 CFR 211.180(e), EU GMP Ch 1	ApqrMetricsResponse
18	AuditInspectionReadinessAgent	Inspection Readiness Auditor	ISO 13485:8.2.4, CP 7382.850	InspectionReadinessResponse
19	SystemInterfaceMigrationAgent	Data Migration Specialist	GAMP 5 App D6, ALCOA+	DataMigrationResponse
20	SpreadsheetValidationAgent	Spreadsheet Validator	21 CFR Part 11, GAMP 5 Cat 5	SpreadsheetValidationResponse
21	DisasterRecoveryAgent	Business Continuity & DR Lead	21 CFR 11.10(c), Annex 11 Cl 7/16	DisasterRecoveryResponse
22	VendorAssessmentCsaAgent	Cloud & AI Foundation Assessor	ISO 13485:7.4, ISO/IEC 42001	VendorAssessmentResponse
23	DocumentTriageAgent	S3 File Ingestion Classifier	GAMP 5 Category Selection	DocumentTriageResponse

Technical Implementation Details

1. Dependency Injection (GxpAgentDeps)

All agents inherit dependencies defined in src/agents/deps.py:

@dataclass
class GxpAgentDeps:
    pg_pool: asyncpg.Pool
    qdrant_client: QdrantClient
    memgraph_client: MemgraphDriver
    rustfs_client: RustFSClient
    langfuse_tracker: LangfuseTracker
    system_code: Optional[str] = None
    regulatory_mode: str = "operational"

This enables agents to interact with all three databases and storage engines concurrently while recording hierarchical traces in Langfuse.

2. EvidenceGate Integration

Before an agent receives retrieved context chunks from Qdrant or Memgraph, the chunks pass through EvidenceGate:

• Unverified chunks or chunks failing PostgreSQL hash validation are removed.
• Superseded documents are filtered out.
• The agent prompt receives only verified, active regulatory evidence.

3. Structured Pydantic Output Enforcement

Agents return strictly typed Pydantic models. For example, RiskAssessmentAgent returns:

class RiskItem(BaseModel):
    hazard_id: str
    description: str
    severity: int = Field(ge=1, le=5)
    occurrence: int = Field(ge=1, le=5)
    detection: int = Field(ge=1, le=5)
    rpn: int = Field(ge=1, le=125)
    csa_rigor: Literal["FULL_SCRIPTED_TESTING", "LIMITED_SCRIPTED_TESTING", "UNSCRIPTED_EXPLORATORY_TESTING", "VENDOR_AUDIT_LEVERAGING"]
    mitigation_action: str

class RiskAssessmentResponse(BaseModel):
    system_name: str
    gamp_category: int
    critical_thinking_rationale: str
    risk_items: List[RiskItem]
    overall_risk_level: str
    evidence_citations: List[str]

If an LLM produces an invalid calculation or schema violation, Pydantic AI automatically requests self-correction from the model before returning the response.