Example: Multi-Agent Workflow

When you have multiple agents with different responsibilities, AgenticGuard lets you assign different policies and audit each agent independently.

Architecture

  Orchestrator Agent (ag_orch_xxx)
       │
       ├──► Research Agent (ag_research_xxx)
       │         └── reads GitHub, news APIs
       │
       └──► Action Agent (ag_action_xxx)
                 └── creates issues, sends Slack messages

Each agent has its own API key, proxy URL, and policy.

Research agent policy

version: "1"
rules:
  # Read-only: only GET calls allowed
  - tool: "*"
    action: GET
    decision: allow

  - tool: "*"
    action: "*"
    decision: deny
    message: "Research agent is read-only"

Action agent policy

version: "1"
rules:
  # Allow creates with reasoning
  - tool: github
    action: POST
    require_reasoning: true

  - tool: slack
    action: POST
    require_reasoning: true

  # Block everything else
  - tool: "*"
    action: DELETE
    decision: deny

  - tool: "*"
    action: PUT
    decision: deny

Orchestrator code

from agentic_guard import AgenticGuard, PolicyViolationError

research_gate = AgenticGuard(api_key="ag_research_xxx", mode="enforce")
action_gate = AgenticGuard(api_key="ag_action_xxx", mode="enforce")

RESEARCH_PROXY = "https://api.agentic-guard.com/proxy/ag_research_xxx/"
ACTION_PROXY = "https://api.agentic-guard.com/proxy/ag_action_xxx/"

import httpx

def research_phase(topic: str) -> dict:
    resp = httpx.get(
        f"{RESEARCH_PROXY}api.github.com/search/repositories",
        params={"q": topic},
    )
    return resp.json()

def action_phase(repo: str, finding: str) -> dict:
    try:
        resp = httpx.post(
            f"{ACTION_PROXY}api.github.com/repos/{repo}/issues",
            json={"title": f"Finding: {finding[:60]}", "body": finding},
            headers={"X-Agent-Reasoning": f"Reporting research finding about {repo}"},
        )
        return resp.json()
    except Exception as e:
        return {"error": str(e)}

Audit log view

The audit log shows events from both agents, tagged by agent_id:

agent         tool    action  decision  reasoning
research-bot  github  GET     allowed   —
research-bot  github  POST    denied    —   ← blocked by policy
action-bot    github  POST    allowed   "Reporting research finding"
action-bot    github  DELETE  denied    —   ← blocked by policy

Benefits

• Each agent's blast radius is limited by its own policy
• Any policy violation surfaces immediately in the audit log
• Credentials are shared at the org level — store once, used by all agents