managed-agents 0.2.0

Nomotic governance for AI agents. Works with Claude, LangChain, CrewAI, AutoGen, OpenAI, or any custom framework.

managed-agents

Nomotic governance for AI agents. Works with any framework.

Govern every tool call before it reaches execution. DENY verdicts block the action. ESCALATE routes to human review. Every evaluation is recorded in your Nomotic audit trail.

Install

pip install managed-agents

Quick Start

from managed_agents import NomoticHarness, GovernanceDenied

harness = NomoticHarness(
    api_key="nm_live_...",    # from amp.nomotic.ai/settings
    agent_id="nmc-...",       # from amp.nomotic.ai/identity
    platform="langchain",     # tags evaluations in your audit trail
)

async def execute_tool(tool_name: str, tool_input: dict) -> str:
    await harness.govern(tool_name, tool_input)
    return await your_runtime.execute(tool_name, tool_input)

try:
    await execute_tool("write_file", {"path": "/etc/passwd", "content": "..."})
except GovernanceDenied as e:
    print(f"Blocked: {e}")

With Prescreening

Block destructive patterns before they reach the governance API:

from managed_agents import NomoticHarness, StandardPrescreener

harness = NomoticHarness(
    api_key="nm_live_...",
    agent_id="nmc-...",
    prescreener=StandardPrescreener(
        block_destructive_shell=True,    # rm -rf, DROP TABLE, mkfs, fork bombs
        block_bulk_operations=True,      # >10 emails, >20 files, >5 calendar items
        escalate_credential_access=True, # .ssh/, .env, /etc/passwd
    ),
)

With Policy Presets

from managed_agents import NomoticHarness, ToolPolicy

harness = NomoticHarness(
    api_key="nm_live_...",
    agent_id="nmc-...",
    policy=ToolPolicy.SHELL,       # trust_threshold=0.85, veto: scope+ethics
    # policy=ToolPolicy.FILESYSTEM # trust_threshold=0.75, veto: reversibility
    # policy=ToolPolicy.EMAIL      # trust_threshold=0.70, veto: data_sensitivity
)

Sync Usage

result = harness.govern_sync("bash", {"command": "ls -la"})

Without an API Key

No api_key → pass-through mode. Prescreener still blocks dangerous patterns. Verdicts logged locally to ~/.nomotic/managed-agents.log.

After 10 evaluations, a nudge suggests registering at amp.nomotic.ai for full governance, audit trail, trust scores, and fleet management.

harness = NomoticHarness(agent_id="local-agent")  # no api_key

Framework Support

platform="claude"      # Claude Managed Agents
platform="langchain"   # LangChain / LangGraph
platform="crewai"      # CrewAI
platform="autogen"     # AutoGen
platform="openai"      # OpenAI SDK / Assistants API
platform="custom"      # Any other framework

Session Correlation

from managed_agents import NomoticHarness, GovernedSession

harness = NomoticHarness(api_key="...", agent_id="...")
session = GovernedSession(session_id="sess_abc123", harness=harness)

verdict = await session.govern_and_record("send_email", {"to": "[e-mail]"})
print(session.governance_summary())

Configuration

Parameter	Default	Description
api_key	None	Nomotic API key. None = pass-through mode
agent_id	required	Nomotic agent ID (nmc-...)
platform	"custom"	Framework tag for audit trail
fail_open	True	Allow on governance service error
block_on_escalate	False	Raise GovernanceEscalated on ESCALATE
timeout	3.0	HTTP timeout in seconds
prescreener	None	Pre-evaluation deterministic checks
policy	None	Governance policy preset
tool_name_map	None	Custom tool name → action verb overrides
circuit_breaker	None	CircuitBreaker for API resilience
local_log	True	Log verdicts to ~/.nomotic/managed-agents.log
suppress_nudge	False	Suppress amp.nomotic.ai registration nudge
on_verdict	None	Callback for every verdict
on_escalate	None	Callback for ESCALATE verdicts
on_deny	None	Callback for DENY verdicts
session_id	None	Session ID for multi-turn correlation
org_context	{}	Additional context injected into evaluations

Links

• Nomotic Platform
• Documentation
• GitHub