Skip to content

Safety Guardrails, MCP & A2A Protocols

This document details the safety guarantees, Model Context Protocol (MCP) tool integration, and Agent-to-Agent (A2A) communication protocols provided by tuluat-guardrails and tuluat-protocols.

1. Safety Guardrails Engine (tuluat-guardrails)

The tuluat-guardrails module implements input sanitization, threat mitigation, and output verification filters applied before and after LLM inference.

graph LR
    UserQuery[User Input / Prompt] --> PII[PiiMaskingFilter]
    PII --> Injection[PromptInjectionFilter]
    Injection --> LLM[LLM Provider / WireMock]
    LLM --> OutputVal[OutputValidationFilter]
    OutputVal --> Response[Verified Agent Response]

1.1 PII Masking Filter (PiiMaskingFilter)

  • Supported Modes: EMAIL, CREDIT_CARD, SSN, PHONE
  • Replacement Token: Customizable per agent (e.g. [REDACTED])
  • Behavior: Uses high-performance regex pattern matching to mask sensitive data before transmitting prompts to external LLM providers.

1.2 Prompt Injection Filter (PromptInjectionFilter)

  • Strategies:
  • BLOCK: Immediately halts execution and throws GuardrailBlockedException if injection heuristics (jailbreak phrases, system instruction overrides) match.
  • SANITIZE: Strips out malicious payload instructions and allows execution with sanitized text.

1.3 Output Validation Filter (OutputValidationFilter)

  • Hard Schema Validation: Validates model output against node-declared JSON Schema specifications.
  • Confidence Scoring: Assigns a confidence score (1.0 for schema-valid output, 0.2 for malformed output) and rejects non-compliant responses.

2. Model Context Protocol (MCP) Client Registry (tuluat-protocols)

The MCP registry enables agents to discover and invoke tools exported by external MCP servers (McpServer CRD).

2.1 Component Specifications

  • McpServer CRD: Defines connection endpoint (sse or stdio), authentication secret reference, and active tool capabilities.
  • McpClientRegistryImpl:
  • Manages active SSE/stdio transports to registered MCP servers.
  • Discovers tool manifests (listTools()) and maps them into dynamic skill definitions.
  • Handles tool invocation (callTool(serverName, toolName, arguments)) with structured error handling.

3. Agent-to-Agent (A2A) Protocol Adapter (tuluat-protocols)

The A2A module enables decentralized discovery and communication between autonomous agents running inside or outside the Kubernetes cluster.

3.1 A2A Card (A2aCard)

  • Exports agent metadata: agentId, capabilities, inputSchema, outputSchema, endpointUrl.
  • Enables automatic agent capability matching in multi-agent workflows.

3.2 Inter-Agent Message Relay (A2aAdapterImpl)

  • Relays messages between agents via HTTP/gRPC endpoints.
  • Supports asynchronous delegation where a leader agent delegates sub-tasks to remote domain-specialist agents.