Architecture

Enverif follows a durable manage → execute → audit loop.

Manage

The agent configuration chooses instructions, a model connection, skills, connectors, limits and capability policy. The tool registry assembles first-party tools, connected connector actions and enabled MCP tools into provider-neutral JSON Schema definitions.

Execute

AgentOrchestrator persists a run and user message, dispatches a queue continuation, requests one model turn, persists assistant/tool calls, evaluates every tool against capability policy, executes allowed steps or pauses for approval, persists tool results, then queues the next turn. MySQL is authoritative throughout.

Audit

Model turns, permission decisions, approval requests and tool completion/failure write hash-linked audit events. Run summaries record token use, optional cost estimates and stop reason.

Extension boundaries

Model providers, connector drivers, skills and MCP servers have separate contracts. A new connector does not modify the agent loop. A new model provider does not modify tools. Skills provide procedural context but cannot elevate runtime permissions.