Problem
Autonomous agents can gain tool access, take actions and interact with systems without sufficient governance.
System concept
A governance layer controlling agent identity, permissions, risk scoring, approvals, quarantine and audit trails, sitting between agents and the tools they're allowed to call.
Architecture
Python, FastAPI, PostgreSQL, SQLAlchemy, Docker Compose.
Human control
- Actions above a risk threshold pause for human approval
- Any agent, job, or tenant can be killed immediately
Safeguards
- Tamper-evident audit trail
- Policy enforcement at the tool-call level
- Quarantine isolates a single agent without a full shutdown
What has been built
Core control-plane logic (identity, permissions, risk scoring, kill switches) plus an operational demo dashboard showing agents, risk scores, audit events and kill-switch controls.
Current status
Built and demonstrated internally. Not yet running in a production agent deployment.
Have a process that still depends on spreadsheets, inboxes and manual follow-ups?
Show us the workflow. We'll map where orchestration could remove friction without removing control.