Agent Runtime¶
The Agent Workbench is the heavy internal execution lane for multi-source work. It is not a generic chatbot. It is a task ledger with policy, trace, artifacts, source boundaries, and backend-owned tools.
Runtime Pieces¶
| Component | Responsibility |
|---|---|
LeafEnterprise/agents/service.py |
Task, trace, artifact persistence. |
LeafEnterprise/agents/policies.py |
Role, lane, source-system, provider, and mutation policy. |
LeafEnterprise/agents/runner.py |
OpenAI Agents SDK execution hook and backend tools. |
LeafEnterprise/agents/sql_guard.py |
Read-only SQL validation before any execution tool can run. |
LeafEnterprise/agents/dispatcher.py |
Service Bus dispatch seam for AKS workers. |
scripts/run_agent_service_bus_worker.py |
First worker entrypoint for agent-tasks. |
Design Rule¶
Agents can reason across governed evidence, but deterministic LeafEnterprise calculators and validators remain authoritative.