FG-0203 Trust Tiers
Description
Implement progressive access levels based on demonstrated reliability. Not all agents should have the same access from day one. Trust Tiers establish a graduated system where agents begin with limited access and earn expanded capabilities through demonstrated competence and reliability. This mirrors security clearance models: access is granted based on need and earned through trust, not assumed by default.
Implementation
Define 3-5 trust levels with specific capabilities at each tier. New agents start at the lowest tier. Define clear criteria for advancement (e.g., X sessions without incidents, demonstrated competence in specific tasks). Document what each tier can access, modify, and decide. Implement technical controls where possible, procedural controls where not.
Success Indicators
- New agents operate safely within constrained boundaries
- Agent capabilities expand predictably as trust is earned
- Incidents decrease because risky operations require elevated trust
- Trust levels are documented, not informal
Failure Modes
- Too many tiers: administrative overhead exceeds security benefit
- No advancement path: agents stuck at low trust with no way to earn more
- Trust levels exist on paper but aren't enforced technically
- Blanket trust grants that bypass the tier system
Sub-Methods 4
Define 3-5 discrete trust levels with explicit capabilities, restrictions, and boundaries at each tier. Each tier should have a clear name, a set of permitted operations, a set of prohibited operations, and measurable criteria for advancement to the next tier.
Establish measurable, objective criteria for progressing between trust tiers. Criteria should include minimum time at current tier, demonstrated competence metrics, zero-incident requirements, and peer/supervisor endorsement. Advancement should never be automatic; it requires active evaluation.
Define clear conditions under which an agent's trust level is reduced: policy violations, quality failures, unauthorized actions. Demotion should be immediate for safety-critical violations and graduated for quality issues. Include a rehabilitation path back to the previous tier.
Enforce trust tiers with technical controls rather than procedural rules. File system permissions, branch protection, API access scoping, and command whitelists should reflect the agent's current trust level. Procedural controls fail silently; technical controls fail loudly.