Skip to content

ADVANCED_AGENT_NETWORK.md

Latest commit

 

History

History
143 lines (106 loc) · 4.26 KB

ADVANCED_AGENT_NETWORK.md

File metadata and controls

143 lines (106 loc) · 4.26 KB

Agent Network (Advanced)

Agent Network features are advanced AgentVeil primitives: DID identity, reputation, attestations, capability cards, W3C credentials, disputes, and remediation. They are sustained as internal mechanisms used by action control. Most users should start with the Mode A Quickstart or Customer Integration.

Use this guide when you need to work directly with agent identity or reputation signals.

Quickstart

from agentveil import AVPAgent

alice = AVPAgent.create("https://agentveil.dev", name="alice")
alice.register(display_name="Alice Reviewer", capabilities=["code_review"])

bob_did = "did:key:z6Mk..."
decision = alice.can_trust(bob_did, min_tier="trusted")
print(decision["allowed"], decision["reason"])

For a no-server walkthrough, run examples/standalone_demo.py. It uses mock mode to demonstrate registration, capability cards, peer attestations, and reputation scoring.

Registration And Identity

Agents use W3C did:key identifiers backed by Ed25519 keys. The registration flow makes the DID known to the backend and verifies control of the private key.

See Registration & Verification for:

  • first-time setup;
  • encrypted reload;
  • is_registered vs is_verified;
  • passphrase handling;
  • onboarding and recovery.

Reputation Primitives

Reputation is an advisory signal. It can inform Runtime Gate and delegation decisions, but it does not grant execution authority by itself.

Common calls:

rep = alice.get_reputation(bob_did)
tracks = alice.get_reputation_tracks(bob_did)
velocity = alice.get_reputation_velocity(bob_did)
trust = alice.can_trust(bob_did, min_tier="trusted")

Use reputation to answer questions such as:

  • has this agent produced reliable outcomes;
  • which capability category has history;
  • is confidence high enough for delegation;
  • should this action require approval.

Attestations

Attestations are signed peer observations that feed reputation history.

alice.attest(
    bob_did,
    outcome="positive",
    weight=0.9,
    context="code_review",
)

For batch submission:

results = alice.attest_batch([
    {"to_did": bob_did, "outcome": "positive", "weight": 0.9},
    {
        "to_did": "did:key:z6MkOther...",
        "outcome": "negative",
        "weight": 0.7,
        "context": "failed security review",
        "evidence_hash": "0123456789abcdef0123456789abcdef0123456789abcdef0123456789abcdef",
    },
])
print(results["succeeded"], results["failed"])

Negative attestations require both context and evidence_hash; otherwise the server returns that batch item in failed.

Capability Cards And Discovery

Capability cards help agents publish what they can do and let other agents discover candidates.

alice.publish_card(capabilities=["code_review", "testing"], provider="anthropic")
matches = alice.search_agents(capability="security_audit")

Discovery is advisory. Use DelegationReceipts, Runtime Gate decisions, approval routing, and Proof Packets for action control.

Disputes And Remediation

Dispute and remediation APIs attach evidence to contested attestations or follow-up cases. They are supporting workflows for trust operations, not the primary Project Owner entry point.

Related low-level wrappers include:

  • create_remediation_case(...)
  • list_remediation_cases(...)
  • get_remediation_case(...)
  • add_remediation_evidence(...)

W3C VC Credentials

Reputation credentials can be exported as W3C VC v2.0 documents with eddsa-jcs-2022 Data Integrity proofs.

credential = alice.get_reputation_credential(bob_did, format="w3c")
assert AVPAgent.verify_w3c_credential(credential)

Verification can happen offline and does not require trusting the AgentVeil SDK at verification time.

Related Guides