CLean-agents | build-legal

Agents

pipeline

Pattern

crewai

Framework

$0.1940

Cost / Request

95%

Security Score

✓

Human-in-Loop

Architecture

Agents

Security

Cost Analysis

Decisions

Compliance

graph TD INPUT(["User Input"]) orchestrator{"⚙ orchestrator"} INPUT --> orchestrator doc_analyzer["doc_analyzer"] orchestrator --> doc_analyzer risk_evaluator["risk_evaluator"] orchestrator --> risk_evaluator guardian[/"🛡 guardian"/] INPUT -.->|filter| guardian guardian -.-> orchestrator OUTPUT(["Response"]) orchestrator --> OUTPUT style INPUT fill:#1f6feb,stroke:#58a6ff,color:#fff style OUTPUT fill:#1f6feb,stroke:#58a6ff,color:#fff style orchestrator fill:#d29922,stroke:#e3b341,color:#000 style guardian fill:#f85149,stroke:#f85149,color:#fff

Architecture Summary

System Type	pipeline
Pattern	pipeline
Framework	crewai
Scale	medium
Autonomy Level	L3
Infrastructure	Vector DB: pinecone \| Graph DB: neo4j \| Queue: redis \| Observability: langfuse

orchestrator orchestrator

▼

RolePlan decomposition, task routing, result synthesis

Modelclaude-opus-4-6

Reasoningreact

HITLpre-action

Token Budget2,000

MemoryShort-term

Input Guardsinjection_detection, encoding_detection

Output Guardsschema_validation

Metricsrouting_accuracy: 0.95, recovery_rate: 0.8

doc_analyzer specialist

▼

RoleExtract clauses, obligations, and terms from legal documents

Modelclaude-sonnet-4-6

Reasoningreact

HITLnone

Token Budget4,000

MemoryShort-term, GraphRAG

Input GuardsNone

Output Guardsschema_validation, confidence_threshold

Metricsextraction_accuracy: 0.92

risk_evaluator specialist

▼

RoleScore and categorize contractual risks by severity

Modelclaude-sonnet-4-6

Reasoningreflection

HITLpost-action

Token Budget3,000

MemoryShort-term, GraphRAG

Input GuardsNone

Output Guardsschema_validation

Metricsrisk_classification_accuracy: 0.9

guardian guardian

▼

RoleInput/output safety filtering, PII detection, compliance checks

Modelclaude-haiku-4-5

Reasoningreact

HITLnone

Token Budget500

MemoryShort-term

Input Guardsinjection_detection, encoding_detection, pii_detection

Output Guardspii_masking, content_filter

Metricsdetection_rate: 0.99, false_positive_rate: 0.05

Security Score

80/84 checks passed

Findings by Category

Security Findings

Status	Category	Agent	Check	Detail
✗	ATK-1	doc_analyzer	system_prompt_isolation	Agent 'doc_analyzer' missing input injection detection
✗	ATK-1	doc_analyzer	input_sanitization	Agent 'doc_analyzer' missing input injection detection
⚠	ATK-1	doc_analyzer	instruction_hierarchy	Consider HITL for sensitive operations
✗	ATK-1	risk_evaluator	system_prompt_isolation	Agent 'risk_evaluator' missing input injection detection
✗	ATK-1	risk_evaluator	input_sanitization	Agent 'risk_evaluator' missing input injection detection
⚠	ATK-1	guardian	instruction_hierarchy	Consider HITL for sensitive operations
⚠	ATK-2	doc_analyzer	role_boundary_enforcement	Consider HITL for sensitive operations
⚠	ATK-2	doc_analyzer	multi_turn_tracking	Consider HITL for sensitive operations
⚠	ATK-2	guardian	role_boundary_enforcement	Consider HITL for sensitive operations
⚠	ATK-2	guardian	multi_turn_tracking	Consider HITL for sensitive operations
⚠	ATK-3	orchestrator	system_prompt_protection	Ensure system prompts are not exposed in tool outputs
⚠	ATK-3	doc_analyzer	system_prompt_protection	Ensure system prompts are not exposed in tool outputs
⚠	ATK-3	doc_analyzer	rag_access_control	RAG enabled — ensure access control on retrieval
⚠	ATK-3	risk_evaluator	system_prompt_protection	Ensure system prompts are not exposed in tool outputs
⚠	ATK-3	risk_evaluator	rag_access_control	RAG enabled — ensure access control on retrieval
⚠	ATK-3	guardian	system_prompt_protection	Ensure system prompts are not exposed in tool outputs
⚠	ATK-4	doc_analyzer	inter_agent_auth	Verify agent-to-agent auth at framework level
⚠	ATK-4	doc_analyzer	message_validation	Add input validation for inter-agent messages
⚠	ATK-4	doc_analyzer	privilege_boundaries	Add input validation for inter-agent messages
⚠	ATK-4	risk_evaluator	inter_agent_auth	Verify agent-to-agent auth at framework level
⚠	ATK-4	risk_evaluator	message_validation	Add input validation for inter-agent messages
⚠	ATK-4	risk_evaluator	privilege_boundaries	Add input validation for inter-agent messages
⚠	ATK-4	guardian	inter_agent_auth	Verify agent-to-agent auth at framework level
⚠	ATK-6	orchestrator	recursion_depth_limits	Recursion limits should be configured at framework level
⚠	ATK-6	doc_analyzer	recursion_depth_limits	Recursion limits should be configured at framework level
⚠	ATK-6	risk_evaluator	recursion_depth_limits	Recursion limits should be configured at framework level
⚠	ATK-6	guardian	recursion_depth_limits	Recursion limits should be configured at framework level
⚠	ATK-7	orchestrator	pii_detection	Agent 'orchestrator' has no PII detection on input
⚠	ATK-7	orchestrator	context_isolation	Verify context isolation between users at runtime
⚠	ATK-7	doc_analyzer	pii_detection	Agent 'doc_analyzer' has no PII detection on input
⚠	ATK-7	doc_analyzer	context_isolation	Verify context isolation between users at runtime
⚠	ATK-7	risk_evaluator	pii_detection	Agent 'risk_evaluator' has no PII detection on input
⚠	ATK-7	risk_evaluator	context_isolation	Verify context isolation between users at runtime
⚠	ATK-7	guardian	context_isolation	Verify context isolation between users at runtime

Cost per Agent (per request)

Monthly Projection (10K requests)

LLM Cost	$1940.00
Infrastructure	$200
Total	$2140.00/mo

Optimization Tips

✓ No obvious optimizations needed

D1-Purpose: System type: pipeline

Based on 2 detected responsibilities and deterministic nature.

Alternatives: single-agent, multi-agent, complex-system

D2-Architecture: Pattern: pipeline

2-3 sequential responsibilities with clear data handoffs. Pipeline is simplest.

GDPR

Requirement	Component	Status
Art. 13-14 Transparency	guardrails.output → explainability	Review
Art. 15 Right of access	audit_trail → logging	Configured
Art. 17 Right to erasure	memory → deletion endpoint	Review
Art. 25 Data protection by design	guardrails.input → pii_detection	Configured
Art. 35 DPIA	blueprint.decisions	Review

⚙️ build-legal

Architecture Summary

orchestrator orchestrator

doc_analyzer specialist

risk_evaluator specialist

guardian guardian

Security Score

Findings by Category

Security Findings

Cost per Agent (per request)

Monthly Projection (10K requests)

Optimization Tips

D1-Purpose: System type: pipeline

D2-Architecture: Pattern: pipeline

GDPR