Claude Managed Agents Integration

Anthropic’s Claude Managed Agents platform handles execution infrastructure for Claude-powered agents. The managed-agents package handles governance — what those agents are allowed to do once they’re running. The two work together seamlessly.

Installation

pip install managed-agents

Basic integration

Create a harness with platform="claude" and govern every tool call before it is executed by Claude’s runtime:

python

from managed_agents import NomoticHarness, GovernanceDenied

harness = NomoticHarness(
    api_key="nm_live_...",
    agent_id="nmc-...",
    platform="claude",
)

async def execute_tool(tool_name: str, tool_input: dict) -> dict:
    try:
        await harness.govern(tool_name, tool_input)
        return await claude_runtime.execute(tool_name, tool_input)
    except GovernanceDenied as e:
        return {
            "type": "error",
            "error": {"type": "governance_denied", "message": str(e)}
        }

Handling escalations

By default, escalated actions are routed to your Nomotic approval queue and execution continues if fail_open=True. For Claude agents where you want to pause execution pending human review:

python

harness = NomoticHarness(
    api_key="nm_live_...",
    agent_id="nmc-...",
    platform="claude",
    block_on_escalate=True,
)

With block_on_escalate=True, an escalated tool call raises GovernanceDenied. Your agent can surface this to the user and wait before retrying once the escalation is resolved.

What Nomotic adds

Running on Claude Managed Agents gives you execution infrastructure. Adding managed-agents gives you a cryptographic certificate for your agent, an immutable audit trail of every tool call, a trust score reflecting behavioral history, drift detection if behavior changes over time, and compliance documentation for regulated industries.

Data handling

Nomotic only receives what you pass to harness.govern() — the tool name and input parameters. The full content of Claude’s responses and reasoning stays within Anthropic’s infrastructure.