Governance
Human-in-the-middle. Agents propose, you decide.
The Principle
Human-in-the-middle. Every task that changes something in your vault, your projects, or the outside world requires your explicit approval unless it has been marked as auto-approved in the Approval Matrix.
Agents propose. You decide. This is the foundation of the entire system. It means you never wake up to surprises you did not ask for. It also means you stay in control without having to micromanage every step.
Approval Matrix
The Approval Matrix defines what agents can do on their own and what needs your sign-off. Two categories.
| Category | Actions |
|---|---|
| Auto-Execute | Process daily captures, nightly Chronicle entries, weekly and monthly reflections, task proposals to backlog, content drafts, /distribute |
| Requires Approval | Task execution, skill and workflow modifications, publishing externally, file deletion, external API calls (except gh api for repo monitoring), schedule changes |
Agent Permissions
Each agent has scoped skills and boundaries defined in their workspace. They can only use the skills assigned to them and can only write to the folders their role permits.
Permissions are defined per-agent in their AGENT.md file. This includes which skills they can invoke, which folders they can write to, and what actions trigger an escalation.
See each agent's page for their specific permission boundaries: Cortex, Loom, Radar, Hippocampus, Signal, Sentinel, Axon.
Escalation Rules
There are situations where agents must stop and ask, regardless of their permissions. These are non-negotiable.
- Any action with medium or higher risk to data, reputation, or finances
- Claims that cannot be verified with evidence
- Deployments to production environments
- Changes to account settings, billing, or credentials
- Sensitive topics involving personal information, family, or health
Rejection Log
The Rejection Log tracks patterns of what you have said no to. When an agent proposes a task, it checks the log first. If the proposal matches a previously rejected pattern, the agent either skips it entirely or adjusts the approach.
This is how the system learns what you do not want. Over time, agents stop proposing things that waste your review time. The log is a living document that gets smarter with every rejection.
How Trust Grows
The system starts restrictive on purpose. Every new action type requires approval until you are comfortable with how the agent handles it.
As trust builds, you expand the auto-execute list. An action that needed approval last month might run automatically this month because the agent has proven it handles it well. You are always the one who updates the matrix. The system never promotes itself.