Agent Runtime

# AgentMesh Runtime **Execution supervisor for multi-agent sessions — privilege rings, saga orchestration, and governance enforcement** *Part of the [Agent Governance Toolkit](https://github.com/microsoft/agent-governance-toolkit)* [![CI](https://github.com/microsoft/agent-governance-toolkit/actions/workflows/ci.yml/badge.svg)](https://github.com/microsoft/agent-governance-toolkit/actions/workflows/ci.yml) [![License](https://img.shields.io/badge/license-MIT-blue.svg)](../../LICENSE) [![Python](https://img.shields.io/badge/python-3.11+-blue.svg)](https://python.org) [![PyPI](https://img.shields.io/pypi/v/agentmesh-runtime)](https://pypi.org/project/agentmesh-runtime/)

[!IMPORTANT] Public Preview — The agentmesh-runtime package on PyPI is a Microsoft-signed public preview release. APIs may change before GA.

Note: This package was renamed from agent-runtime to agentmesh-runtime to avoid a PyPI name collision with the AutoGen team's package. The agent-hypervisor package remains the canonical upstream implementation; agentmesh-runtime is a thin re-export wrapper for incremental import migration.

What is Agent Runtime?¶

Agent Runtime provides execution-level supervision for autonomous AI agents. While Agent OS handles policy decisions and AgentMesh handles trust/identity, Agent Runtime enforces those decisions at the session level:

Execution Rings — 4-tier privilege model (Ring 0–3) controlling what agents can do at runtime
Shared Sessions — Multi-agent session management with consistency modes (strict, eventual, causal)
Saga Orchestration — Compensating transactions for multi-step agent workflows
Kill Switch — Immediate termination with audit trail and blast radius containment
Joint Liability — Attribution tracking across multi-agent collaborations
Audit Trails — Hash-chained, append-only execution logs

Quick Start¶

pip install agentmesh-runtime

from hypervisor import Hypervisor, SessionConfig, ConsistencyMode

# Create the runtime supervisor
hv = Hypervisor()

# Create a governed session
session = await hv.create_session(
    config=SessionConfig(consistency_mode=ConsistencyMode.EVENTUAL)
)

# Execute with privilege enforcement
result = await session.execute(
    agent_id="researcher-1",
    action="tool_call",
    tool="web_search",
    ring=2  # restricted privilege ring
)

Architecture¶

┌─────────────────────────────────────────────────────────────────────┐
│                        Agent Runtime                                 │
├─────────────┬──────────────────┬──────────────────┬─────────────────┤
│  Execution  │     Session      │      Saga        │    Liability    │
│   Rings     │   Management     │  Orchestration   │    Tracking     │
│             │                  │                  │                 │
│  Ring 0:    │  Create/join     │  Multi-step      │  Attribution    │
│   System    │  Consistency     │  Compensation    │  Vouching       │
│  Ring 1:    │  Checkpoints     │  Rollback        │  Slashing       │
│   Trusted   │  Merge/fork      │  Recovery        │  Quarantine     │
│  Ring 2:    │                  │                  │                 │
│   Standard  │                  │                  │                 │
│  Ring 3:    │                  │                  │                 │
│   Sandboxed │                  │                  │                 │
└─────────────┴──────────────────┴──────────────────┴─────────────────┘

Ecosystem¶

Agent Runtime is one of 7 packages in the Agent Governance Toolkit:

Package	Role
Agent OS	Policy engine — deterministic action evaluation
AgentMesh	Trust infrastructure — identity, credentials, protocol bridges
AgentMesh Runtime	Execution supervisor — rings, sessions, sagas (this package)
Agent SRE	Reliability — SLOs, circuit breakers, chaos testing
Agent Compliance	Regulatory compliance — GDPR, HIPAA, SOX frameworks
Agent Marketplace	Plugin lifecycle — discover, install, verify, sign
Agent Lightning	RL training governance — governed runners, policy rewards

License¶

MIT — see LICENSE.