Agent Trust Protocol

Manage and update agent trust scores with Bayesian updates, domain-specific trust, revocation, forgetting, and visualize trust via dashboard.

Agent Trust Protocol (ATP)

Establish, verify, and maintain trust between AI agents. Bayesian trust scoring with domain-specific trust, revocation, forgetting curves, and a visual dashboard.

Install

bash

git clone https://github.com/FELMONON/trust-protocol.git
# No dependencies beyond Python 3.8+ stdlib
# Pair with skillsign for identity: https://github.com/FELMONON/skillsign

Quick Start

bash

# Add an agent to your trust graph
python3 atp.py trust add alpha --fingerprint "abc123" --score 0.7

# Record interactions — trust evolves via Bayesian updates
python3 atp.py interact alpha positive --note "Delivered clean code"
python3 atp.py interact alpha positive --domain code --note "Tests passing"

# Check trust
python3 atp.py trust score alpha
python3 atp.py trust domains alpha

# View the full graph
python3 atp.py status
python3 atp.py graph export --format json

# Run the full-stack demo (identity → trust → dashboard)
python3 demo.py --serve

Commands

Trust Management

bash

atp.py trust add <agent> --fingerprint <fp> [--domain <d>] [--score <0-1>]
atp.py trust list
atp.py trust score <agent>
atp.py trust remove <agent>
atp.py trust revoke <agent> [--reason <reason>]
atp.py trust restore <agent> [--score <0-1>]
atp.py trust domains <agent>

Interactions

bash

atp.py interact <agent> <positive|negative> [--domain <d>] [--note <note>]

Challenge-Response

bash

atp.py challenge create <agent>
atp.py challenge respond <challenge_file>
atp.py challenge verify <response_file>

Graph

bash

atp.py graph show
atp.py graph path <from> <to>
atp.py graph export [--format json|dot]
atp.py status

Dashboard

bash

python3 serve_dashboard.py          # localhost:8420
python3 demo.py --serve             # full demo + dashboard

Moltbook Integration

bash

python3 moltbook_trust.py verify <agent>    # check agent trust via Moltbook profile

How Trust Works

• Bayesian updates: Each interaction shifts trust scores with diminishing deltas (prevents thrashing)
• Negativity bias: Negative interactions hit harder than positive ones boost
• Domain-specific: Trust an agent for code but not for security advice
• Forgetting curves: Trust decays without interaction (R = e^(-t/S))
• Revocation: Immediate drop to floor, restorable at reduced score
• Transitive trust: If you trust A and A trusts B, you partially trust B (with decay)

Integration with skillsign

ATP builds on skillsign for identity:

1. Agents generate ed25519 keypairs with skillsign
2. Agents sign skills, others verify signatures
3. Verified agents get added to the ATP trust graph
4. Interactions update trust scores over time

Triggers

"check trust", "trust score", "trust graph", "verify agent", "agent trust", "trust status", "who do I trust", "trust report"