Hunter Agent Skill v0.9

Production-control layer for the Hunter Agent Skills security research framework.

Architecture

                    Hunter Agent Skills
                            │
                    Multi-Agent Graph
                            │
              ┌─────────────┴─────────────┐
              │                           │
         Knowledge/RAG                Agent Runtime
              │                           │
              └─────────────┬─────────────┘
                            ▼
                     Policy-as-Code
                            │
                 ┌──────────┴──────────┐
                 ▼                     ▼
             Scope Guard          Approval Gate
                 │                     │
                 └──────────┬──────────┘
                            ▼
                     Sandbox Boundary
                            │
                 ┌──────────┼──────────┐
                 ▼          ▼          ▼
             Workers      MCP       Evidence
                 │          │          │
                 └──────────┼──────────┘
                            ▼
              PostgreSQL + Graph Store
                            │
                     Observability
                            │
                            ▼
                  Signed Research Bundle

v0.9 production controls

• PostgreSQL event-store adapter contract
• Redis/NATS/Kafka message-bus contract
• Neo4j/ArangoDB graph-store contract
• policy-as-code engine
• sandbox execution boundary
• OpenTelemetry-compatible tracing contract
• Prometheus-compatible metrics contract
• Ed25519 signing contract
• benchmark corpus and evaluation plan
• OpenAPI control-plane/dashboard contract

Security boundary

The order is intentional:

Agent decision
    ↓
Policy evaluation
    ↓
Scope validation
    ↓
Approval if required
    ↓
Sandbox/tool boundary
    ↓
Execution
    ↓
Evidence

An agent cannot bypass policy by directly invoking an infrastructure adapter.

Production migration

Reference adapters are dependency-light contracts. Bind them to:

• PostgreSQL + psycopg/asyncpg
• Redis/NATS/Kafka
• Neo4j/ArangoDB
• OpenTelemetry
• Prometheus
• Ed25519 implementation
• container/VM sandbox
• selected MCP runtime

without changing the Agent Skill interfaces.

v1.0 target

• stable Agent Skills package specification
• production MCP implementation
• hardened sandbox
• full persistence migrations
• graph-based knowledge/hypothesis engine
• approval UI
• benchmark CI gates
• signed reproducibility bundles
• OpenTelemetry + Prometheus dashboards
• provider-agnostic LLM routing
• end-to-end integration tests