// Click 'Run Multi-Agent Mesh' to observe live DAG orchestration, PII tokenization, statutory verification, and HITL interruption...
Type or paste text containing Brazilian CPF, US SSN, Canadian SIN, or monetary compensation. The gateway replaces sensitive identifiers with cryptographic surrogate tokens before entering external model context.
// Tokenized payload will appear here...
// Restored plaintext will appear here...
Simulate Tier 3 red-teaming attacks against the agent mesh perimeter. Freeform untrusted inputs (e.g. manager calibration justifications, free-text override requests) are inspected for prompt injections, system delimiter tampering, and canary token exfiltration before entering agent context.
// Delimiter sandboxing output will appear here...
// Security assessment snapshot...
Expose verified compensation engines, Brazilian CLT compliance verifiers, and Zero-Retention PII vaults directly to AI coding agents (Claude Desktop, Cursor, Antigravity) via standardized JSON-RPC 2.0 over Stdio and HTTP/SSE transports.
// Click 'Dispatch JSON-RPC Frame' to test execution...
{
"mcpServers": {
"people-mesh": {
"command": "people-mesh",
"args": ["--mcp"]
}
}
}
// Click 'Execute CI Golden Benchmarks' to evaluate edge cases across Brazil, US, and Canada...
In high-stakes enterprise engineering calibration, promotions and merit exceptions are reviewed by cross-functional panels. This panel convenes an Advocate (Sponsor), a Skeptic (Bar-Raiser), and an Equity & Budget Auditor, moderated by a Consensus Moderator with a formal Reflexion loop that catches vague coaching and self-corrects into measurable quarterly OKRs.
In high-stakes enterprise People Operations, human calibration approvals take hours or days. This durable engine persists an append-only, tamper-evident event stream to SQLite Write-Ahead Logging (WAL). Simulate a sudden container kill/pod eviction, recover state instantly with 0.0% data loss without re-running LLM inferences, and observe automated backward Saga compensation rollbacks on rejection.
ADR-001: Stateful Graph Orchestration vs. Autonomous ReAct Loops
Why high-stakes enterprise HR requires deterministic state machines with JSON snapshot checkpointing over open-ended ReAct loops.
ADR-002: Zero-Retention Privacy Gateway & PII Tokenization
Perimeter ingress surrogate tokenization for Brazil LGPD, US FLSA, and Canadian PIPEDA with cryptographic shredding.
ADR-003: Adversarial Prompt Injection Defense & Canary Tripwires
Multi-vector heuristic threat scoring, XML delimiter sandboxing, Canary tripwire assertion gates, and CI regression blocking (Tier 3 evals).
ADR-004: Standardizing on Model Context Protocol (MCP)
Exposing deterministic People operations tools as standardized MCP JSON-RPC 2.0 endpoints over stdio and HTTP/SSE. See MCP Server Integration Guide.
ADR-005: LLM-as-a-Judge Semantic Rubrics & Demographic Parity Auditing
Semantic evaluation framework for Faithfulness, Constructive Executive Tone, and Counterfactual Statistical Parity (Δ ≤ 0.0001).
ADR-006: Multi-Agent Calibration Committee Deliberation & Reflexion Loops
Dialectic debate swarms (Advocate, Skeptic, Equity Auditor, Moderator) with verbal reinforcement and OKR self-correction.
ADR-007: Durable Execution Engine, Event Sourcing & Backward Saga Rollbacks
Append-only event streams, SQLite WAL persistence, zero-loss crash recovery without LLM re-execution, backward saga compensations, and Temporal / Trigger.dev adapters.
RFC-001: Enterprise Agent Standards & Tool Contract Specification
Formal guidelines for AgentSpec, ToolSpec, sliding-window circuit breakers, and CI eval merge gates.