ADR 006: Agent Runtime with EvalView Guardrails#

Status#

Accepted

Date#

2024-04-01

Context#

Mutx deploys autonomous agents that execute actions in customer environments. We need robust safety guardrails.

Decision#

Implement a layered agent runtime:

1. Agent Runtime Service: Manages agent lifecycle, tools, and execution
2. EvalView Guardrail: Hypervisor-level security layer with local LLM judge
3. Self-Healing Service: Monitors and recovers failed agent runs

Consequences#

Positive#

• Safety: Local LLM judge evaluates actions before execution
• Observability: Full audit trail of agent decisions
• Resilience: Automatic recovery from failures
• Extensible: Support for multiple agent frameworks (LangChain, OpenClaw, n8n)

Negative#

• Latency: Guardrail checking adds execution overhead
• Complexity: Multiple services to manage and debug

Alternatives Considered#

• No guardrails: Rejected - unacceptable security risk
• External audit only: Rejected - too slow for autonomous agents
• Rule-based filtering: Rejected - insufficient for complex agent actions

References#

• Agent Runtime Documentation
• Security Architecture