kanoniv-autogen

Cryptographic identity and delegation for AutoGen agents.

Every agent gets a did:agent: DID. Authority flows from human to agent to sub-agent, narrowing at each step. Revocable at any point.

Install

pip install kanoniv-autogen

Quick Start

from kanoniv_agent_auth import AgentKeyPair
from kanoniv_autogen import AuthorityManager, DelegatedAgent

# Root authority
human = AgentKeyPair.generate()
authority = AuthorityManager(human)

# Authorize agents with scoped permissions
researcher = DelegatedAgent("researcher", actions=["search", "summarize"], max_cost=5.0)
writer = DelegatedAgent("writer", actions=["write"], max_cost=3.0)
authority.authorize(researcher)
authority.authorize(writer)

# Verified execution
researcher.act("search", {"query": "AI safety", "cost": 0.50})
# Returns (invoker_did, root_did, chain, depth)

# Blocked: wrong scope
writer.act("search", {"query": "not allowed", "cost": 0.10})
# Raises ValueError: action 'search' not in scope

Sub-Delegation

helper = DelegatedAgent("helper")
researcher.delegate_to(helper, actions=["search"], max_cost=2.0)
# helper can search at max $2, inherits researcher's other caveats

Revocation

authority.revoke_agent(writer)
# All subsequent writer.act() calls raise ValueError

Audit Trail

for entry in authority.audit_report():
    print(f"{entry['agent']} ({entry['did'][:20]}...) -> {entry['action']}")

API

AuthorityManager

Method	Description
authorize(agent)	Delegate from root to agent
revoke_agent(agent)	Revoke delegation
is_revoked(hash)	Check revocation status
audit_report()	Combined audit log

DelegatedAgent

Property/Method	Description
did	Agent's did:agent: DID
act(action, args)	Execute with verified delegation
delegate_to(other, actions, max_cost)	Sub-delegate

Links

• kanoniv-agent-auth - Core identity and delegation library
• AutoGen - Multi-agent conversation framework
• MCP Auth Proposal

License

MIT