Skip to content

Latest commit

 

History

History
50 lines (40 loc) · 4.07 KB

README.md

File metadata and controls

50 lines (40 loc) · 4.07 KB

A2A x Iron Book — Least-Privilege Agent Task Handoff (Extension Demo)

Iron Book SDK target: ironbook-sdk >= 0.3.2.
Docs (Iron Book SDK): https://pypi.org/project/ironbook-sdk — Quick Start, methods, calls, data types.
Iron Book comprehensive solution overview: https://docs.identitymachines.com.
Docs (Extension): Iron Book A2A Extension page https://pypi.org/project/ironbook-a2a-extension.
This demo is governance-focused; no real LLM call is made (document for the Summarizer agent action is simulated).

This repository demonstrates a profile/data A2A extension (x-ironbook) that adds zero-trust, policy-gated handoffs between agents:

  • Least-privilege delegation: a Triage agent (requester) can only delegate; a Summarizer (executor) agent is the only one allowed to infer. Capabilities are enforced by policy, not hope.
  • Two-decision guardrail: we separately verify that the requester is allowed to ask and the executor is allowed to act, solving the classic confused-deputy problem.
  • Enterprise guardrails in Rego: model allow-list, region & data-class, PII=no, budget caps, etc.; all evaluated server-side with an injected behavioral 0–100 trust score.

If you operate AI agents, this pattern is a fast path to governed, auditable agent-to-agent workflows. PRs & feedback welcome!

Demo Rego Policy (Rego v1) Included

policies/llm_guard.rego enforces:

  • Requester must have the delegate capability and sufficient trust.
  • Executor must have openai_infer, allowed model/region/data_classification, no PII, and budget within limits.

What Happens on Run

  1. Triage agent (requester) registers in Iron Book's secure agent registry with delegate capability only
  2. Summarizer agent (executor) registers in Iron Book's secure agent registry with openai_infer capability
  3. Triage sends an inference (mocked) action delegation request to Summarizer with the Iron Book extension activated using a secure one-shot Iron Book token
  4. Two policy decisions are made:
    • Requester check: Triage has a valid one-shot token, delegate + sufficient trust (all Iron Book agents have a 0-100 trust score)
    • Executor check: Summarizer has has a valid one-shot token, openai_infer + meets guardrails (defined in the provided policy)
  5. If all checks pass, the handoff succeeds (demo result returned)

Two-Decision Pattern

  1. Requester decision: Summarizer calls Iron Book's policy_decision() with Triage DID + Triage one-shot token and context.role="requester".
  2. Executor decision: Summarizer calls Iron Book's policy_decision() with Summarizer DID + Summarizer one-shot token and context.role="executor".

Notes On A2A Compliance

  • Declaration: Summarizer advertises the extension in its AgentCard (/agent-card) under capabilities.extensions[] with a versioned URI.
  • Activation: Triage sets X-A2A-Extensions: <URI> on the request. Summarizer echoes the header on success.
  • Data placement: All extension data are placed under params.metadata["<URI>/…"] (namespaced keys). No core schema changes.

Troubleshooting

  • Module errors: Use python -m agents.summarizer instead of python agents/summarizer.py, and python -m agents.requester instead of python agents/requester.py
  • API key issues: Get your Iron Book API key at https://ironbook.identitymachines.com
  • Agent Not Found: Your agent's DID is be stripped of all non-alphanumeric characters (if you name your Iron Book agent "a2a-summarizer", your DID will be 'did:web:agents.identitymachines.com:a2asummarizer')
  • Port conflicts: Change SUMMARIZER_PORT in .env
  • Terminals: Remember, you need to run Summarizer (server-side; run this first) and Triage in separate terminal windows

Customization & Improvements

  • Real LLM Inference: Add real LLM inference and agents (e.g., using ADK)
  • Customize Policy: Add more Trust Score gates, prompt secrets detection, etc.
  • Expand Use Cases: Don't limit yourself to LLM calls - use MCP to gatekeep tool access, and more