Trust Tiers Explained
How MeshGuard uses trust levels to enable graduated autonomy for AI agents.
The Problem: Not All Agents Are Equal
In any organization deploying AI agents, you'll have a spectrum:
- A brand new chatbot that's still being tested
- A customer service agent that's been running for months
- A data analysis agent with access to sensitive databases
- An admin agent that can modify system configurations
Treating all of these the same is either:
- Too permissive: New agents get access they haven't earned
- Too restrictive: Proven agents can't do their jobs
The Solution: Trust Tiers
MeshGuard implements a graduated trust model with four tiers:
┌─────────────────────────────────────────────────────────┐
│ │
│ Privileged ████████████████████████████████ (4) │
│ │
│ Trusted ██████████████████████ (3) │
│ │
│ Verified ████████████████ (2) │
│ │
│ Anonymous ████████ (1) │
│ │
└─────────────────────────────────────────────────────────┘
Increasing permissions and autonomy →Each tier comes with different default permissions, rate limits, and capabilities.
The Four Tiers
1. Anonymous (Lowest Trust)
Who: Unverified agents, public-facing bots, agents in testing.
Default Permissions:
- Read-only access to public data
- No write operations
- No external API calls
- Heavily rate-limited
Use Cases:
- Public chatbots answering FAQs
- Demo agents on marketing sites
- Agents in development/sandbox
Example Policy:
tier: anonymous
default_rules:
- action: "read:public_*"
effect: allow
- action: "*"
effect: deny
reason: "Anonymous agents have read-only access"
rate_limit: 10/minute2. Verified (Basic Trust)
Who: Agents with confirmed identity, passed initial validation.
Default Permissions:
- Read access to most non-sensitive data
- Limited write operations
- Standard rate limits
- Basic tool access
Use Cases:
- Internal chatbots for employees
- Customer service agents (basic tier)
- Data retrieval agents
Example Policy:
tier: verified
default_rules:
- action: "read:*"
effect: allow
- action: "write:notes"
effect: allow
- action: "write:email"
effect: allow
conditions:
- "recipient.internal == true" # Internal only
- action: "write:*"
effect: deny
rate_limit: 100/minute3. Trusted (Elevated Trust)
Who: Agents with proven track record, handling sensitive operations.
Default Permissions:
- Full read access
- Most write operations
- Can delegate to other agents
- Access to sensitive tools
Use Cases:
- Senior customer service agents (refunds, account changes)
- Data processing agents
- Workflow automation agents
Example Policy:
tier: trusted
default_rules:
- action: "read:*"
effect: allow
- action: "write:*"
effect: allow
- action: "delete:*"
effect: deny
reason: "Deletion requires privileged access"
delegation:
enabled: true
max_depth: 2 # Can delegate, delegates cannot re-delegate
rate_limit: 1000/minute4. Privileged (Maximum Trust)
Who: Critical system agents, admin-level operations.
Default Permissions:
- Full access to all operations
- Can modify policies for lower tiers
- Unlimited delegation
- Minimal rate limiting
Use Cases:
- System administration agents
- Security monitoring agents
- Incident response automation
Example Policy:
tier: privileged
default_rules:
- action: "*"
effect: allow
delegation:
enabled: true
max_depth: unlimited
rate_limit: unlimited
audit_level: enhanced # Extra logging for privileged actionsTrust Tier Behavior Matrix
| Capability | Anonymous | Verified | Trusted | Privileged |
|---|---|---|---|---|
| Read public data | ✅ | ✅ | ✅ | ✅ |
| Read internal data | ❌ | ✅ | ✅ | ✅ |
| Read sensitive data | ❌ | ❌ | ✅ | ✅ |
| Write operations | ❌ | 🟡 Limited | ✅ | ✅ |
| Delete operations | ❌ | ❌ | ❌ | ✅ |
| External API calls | ❌ | 🟡 Limited | ✅ | ✅ |
| Delegate to others | ❌ | ❌ | ✅ | ✅ |
| Modify policies | ❌ | ❌ | ❌ | ✅ |
Implementing Trust Tiers
Creating Agents with Specific Tiers
from meshguard import MeshGuardClient
# Admin client for agent management
admin = MeshGuardClient(admin_token="your-admin-token")
# Create a verified agent
verified_agent = admin.create_agent(
name="customer-service-bot",
trust_tier="verified",
tags=["customer-service", "support"],
)
# Create a trusted agent
trusted_agent = admin.create_agent(
name="refund-processor",
trust_tier="trusted",
tags=["finance", "refunds"],
)Checking Permissions Based on Tier
When an agent makes a request, MeshGuard automatically considers its tier:
# Verified agent trying to delete
verified_client = MeshGuardClient(agent_token=verified_agent.token)
decision = verified_client.check("delete:customer_record")
# → DENIED (verified agents can't delete)
# Privileged agent trying to delete
privileged_client = MeshGuardClient(agent_token=privileged_agent.token)
decision = privileged_client.check("delete:customer_record")
# → ALLOWEDTier-Specific Policies
You can write policies that apply to specific tiers:
name: financial-operations
version: 1
rules:
# All tiers can view balances
- action: "read:account_balance"
effect: allow
# Only trusted+ can process refunds
- action: "write:refund"
effect: allow
conditions:
- "agent.trust_tier IN ['trusted', 'privileged']"
- "request.amount <= 500"
# Only privileged can process large refunds
- action: "write:refund"
effect: allow
conditions:
- "agent.trust_tier == 'privileged'"
- "request.amount > 500"Promoting and Demoting Agents
Trust isn't static. Agents can be promoted or demoted based on:
Promotion Criteria
- Time: Agent has operated without issues for X days
- Volume: Successfully handled Y requests
- Review: Human administrator approval
- Certification: Passed automated testing
# Promote an agent after review
admin.update_agent(
agent_id="customer-service-bot",
trust_tier="trusted",
reason="Promoted after 30-day review period",
)Demotion Triggers
- Policy violations: Multiple denied actions
- Anomalies: Unusual behavior patterns
- Incidents: Involvement in a security event
- Manual: Administrator decision
# Demote an agent after incident
admin.update_agent(
agent_id="rogue-bot",
trust_tier="anonymous",
reason="Demoted pending investigation",
)Delegation and Trust Tiers
When agents delegate work to other agents, trust tiers determine what's possible.
The Permission Ceiling Principle
A delegated agent cannot exceed the permissions of its delegator:
Trusted Agent (can: read, write)
└── Delegates to → Verified Agent
└── Effective permissions: read, write
(Ceiling applied: can't exceed Trusted)Tier-Based Delegation Rules
| Delegator Tier | Can Delegate To |
|---|---|
| Anonymous | ❌ Cannot delegate |
| Verified | ❌ Cannot delegate |
| Trusted | Anonymous, Verified |
| Privileged | All tiers |
# Trusted agent delegation policy
tier: trusted
delegation:
enabled: true
allowed_target_tiers: ["anonymous", "verified"]
max_chain_depth: 2
require_audit: trueMonitoring Trust Tiers
Audit Queries by Tier
# Find all actions by verified agents
verified_actions = client.get_audit_log(
filters={"agent.trust_tier": "verified"},
limit=1000,
)
# Find denials for trusted agents (potential issues)
trusted_denials = client.get_audit_log(
filters={
"agent.trust_tier": "trusted",
"decision": "deny",
},
)Tier Distribution Dashboard
Monitor your agent ecosystem:
┌────────────────────────────────────────────┐
│ Agent Distribution │
├────────────────────────────────────────────┤
│ Privileged ██ 2 (4%) │
│ Trusted ████████ 12 (24%) │
│ Verified ████████████████████ 30 (60%) │
│ Anonymous ██████ 6 (12%) │
└────────────────────────────────────────────┘Best Practices
1. Start Restrictive
New agents should always start at verified or lower:
# Default to verified, promote based on behavior
new_agent = admin.create_agent(
name="new-assistant",
trust_tier="verified", # Never start at trusted/privileged
)2. Document Promotion Criteria
Define clear criteria for each tier transition:
| From | To | Criteria |
|---|---|---|
| Anonymous | Verified | Identity confirmed, basic testing passed |
| Verified | Trusted | 30 days clean operation, 1000+ requests, human review |
| Trusted | Privileged | Security review, business justification, exec approval |
3. Audit Privileged Actions
All privileged agent actions should get enhanced logging:
tier: privileged
audit:
level: enhanced
include_context: true
notify_on: ["delete:*", "admin:*"]4. Regular Reviews
Schedule periodic reviews of agent tiers:
- Weekly: Check for anomalies in trusted+ agents
- Monthly: Review agents due for promotion
- Quarterly: Full tier audit
Conclusion
Trust tiers enable graduated autonomy:
- New agents operate with guardrails
- Proven agents earn more capability
- Critical operations require elevated trust
- The principle of least privilege is systematically enforced
This creates a secure, scalable foundation for deploying AI agents at enterprise scale.
Implement Trust Tiers Today
