AIP plugin: cryptographic identity and delegation for ADK agents

[sunilp]

I published an ADK adapter for the Agent Identity Protocol (AIP) that gives every agent in your ADK hierarchy a verifiable cryptographic identity.

What it does

from aip_agents import AIPConfig
from aip_agents.adapters.adk import AIPAdkPlugin

plugin = AIPAdkPlugin(AIPConfig(app_name="my-app", log_tokens=True))
plugin.register(runner)  # walks agent tree recursively

The plugin recursively walks your agent tree and:

• Assigns each agent an Ed25519 keypair and aip:key:ed25519:z... identifier
• Creates delegation chains from parent to sub-agents with scope attenuation
• Provides X-AIP-Token headers for authenticated tool calls

Scope narrows at every delegation hop. A sub-agent can never exceed its parent's permissions.

Why

MCP and A2A lack an identity layer for multi-agent delegation. OAuth covers single-hop auth but not delegation chains. AIP adds verifiable identity that flows across hops, with 0.189ms verification overhead in Python.

Install

pip install aip-agents[adk]

Links

• Tutorial: Add Cryptographic Identity to Your AI Agents in 5 Minutes
  (ADK walkthrough included)
• Source: github.com/sunilp/aip
• Paper: arXiv:2603.24775
• Related PR: feat(a2a): Add skill-level security support to AgentCardBuilder #4895 (skill-level security for ADK)

Would love feedback on how this fits with ADK's security model.

[surajksharma07]

The recursive tree-walk with per-agent Ed25519 keys and scope attenuation at every delegation hop is a really solid way to close the multi-agent identity gap, 0.189ms verification overhead is impressive too. On your question of how this fits ADK's security model: there's a GCP-based agent_identity/CredentialManager integration natively but it's single-hop OAuth token lifecycle not cross-hop delegation like AIP so they're complementary rather than overlapping. Closing this as Show and Tell since it's a third-party plugin, keep sharing, this is genuinely useful work.