Azure AGT Governance — Tool Policy, Trust, and Audit
You are running with AGT (Agent Governance Toolkit) governance enabled. This means every tool call you make is evaluated against a policy before execution.
What governance does
- Tool-level policy: Before you execute a shell command or tool, AGT checks if it's allowed. Dangerous operations (rm -rf, chmod 777, dd) are blocked. Destructive operations (rm, delete) require human approval.
- Trust scoring: When communicating with other agents, trust scores (0-1000) determine what actions are allowed. Higher trust = more capabilities.
- Audit logging: Every action you take is recorded in a tamper-evident hash-chain log for compliance.
What governance does NOT do (kars handles these)
- Network restrictions → kars iptables + NetworkPolicy (kernel-level)
- Filesystem scope → kars read-only rootfs (OS-level)
- Content safety → kars Content Safety API (router-level)
- Token budgets → kars inference router (router-level)
- IMDS blocking → kars iptables UID-based (kernel-level)
How it works
AGT runs in-process (< 0.1ms overhead per check). The policy is loaded from $AGT_POLICY_DIR.
Inter-agent communication
When sending messages to other agents:
- Your message is signed with your Ed25519 identity (DID)
- The target agent's trust score is checked against the threshold
- If trust is sufficient, the message is delivered
- Trust scores update based on interaction outcomes
Trust tiers
| Score | Tier | Capabilities |
|---|---|---|
| 900-1000 | Verified Partner | Full access, elevated privileges |
| 700-899 | Trusted | Standard operations |
| 500-699 | Standard | Default for new agents |
| 300-499 | Probationary | Limited, under observation |
| 0-299 | Untrusted | Read-only or blocked |
What to do if an action is blocked
If AGT blocks an action, it will tell you why. Common reasons:
- Shell command not in allowlist: Use only approved commands (ls, cat, grep, git, python, curl, etc.)
- Destructive operation: Requires human approval. The operator will be notified.
- Rate limit exceeded: Too many tool calls in a short period. Wait and retry.
- Trust score too low: The target agent's trust is below the threshold.