nexus-agentos 1.7.4

Agent Operating System — Production-ready multi-model agent framework with Tool-Using Agent, LLM Provider abstraction (OpenAI/DeepSeek/Anthropic), Function Calling, streaming, retry, checkpoint/resume, A2A protocol, swarm coordination, comprehensive observability, and enterprise features (RBAC, multi-tenancy, audit logging, API key management).

Nexus AgentOS

Production-grade multi-model agent framework. Build autonomous agents that run on anyone's machine — zero config, three providers, full observability.

---

Why AgentOS?

Most agent frameworks force you to wire up providers, retries, guardrails, and observability by hand. AgentOS ships them as first-class citizens — built into the core architecture, not bolted on.

Capability	AgentOS	LangChain	CrewAI	AutoGen
Multi-provider auto-detect	✅ Zero-config	❌ Manual	❌	❌
A2A Protocol (Agent-to-Agent)	✅ Native	❌	❌	❌
MCP Protocol (Tool integration)	✅ Native	✅ External	❌	❌
Memory Pyramid (STM→WM→LTM)	✅ Built-in	❌	❌	❌
HITL (Human-in-the-Loop)	✅ Built-in	❌	❌	❌
Sandbox execution	✅ Process/Docker	❌	❌	❌
Guardrails (PII/toxicity/injection)	✅ 6 built-in	❌	❌	❌
OpenTelemetry bridge	✅ Native	❌	❌	❌
DI Container	✅ Built-in	❌	❌	❌
Streaming (real-time)	✅	✅	❌	❌
Agent Marketplace	✅ Built-in	❌	❌	❌

---

Quick Start

pip install nexus-agentos

1. Configure (recommended) — interactive wizard

agentos init

Guides you through choosing a provider (OpenAI / DeepSeek / Anthropic), pasting your API key, and tests the connection — no manual export needed.

2. Run a task

agentos "列出当前目录的文件"

3. Or try the demo (no API key needed)

agentos demo

Provider auto-detect: AgentOS detects OPENAI_API_KEY, DEEPSEEK_API_KEY, or ANTHROPIC_API_KEY automatically. Run agentos init to set one up in 30 seconds.

---

Architecture

┌──────────────────────────────────────────────────────────┐
│                     CLI / API Server                      │
├──────────────────────────────────────────────────────────┤
│  ToolAgent (autonomous multi-step reasoning)              │
├────────────┬────────────┬─────────────┬─────────────────┤
│  Guardrails│  Sandbox   │   Memory    │  Observability   │
│ (PII/Toxic │ (Process/  │ (STM→WM→LTM)│ (OTel Bridge +   │
│  /Injection│  Docker)   │             │  Cost Analytics) │
├────────────┴────────────┴─────────────┴─────────────────┤
│  LLM Providers: OpenAI │ DeepSeek │ Anthropic            │
│  A2A Protocol · MCP Protocol · Sequential Pipelines       │
│  Sub-agent Orchestration · DI Container · Plugin System  │
└──────────────────────────────────────────────────────────┘

---

Standout Features

1. Provider Auto-Detection & Resiliency

No manual provider wiring. Set an env var, framework auto-detects. Built-in retry with exponential backoff and circuit breaker.

from agentos.llm import create_provider

# Auto-detect from env
provider = create_provider()  # reads OPENAI_API_KEY → DeepSeek → Anthropic → mock fallback

# Or explicit
provider = create_provider("anthropic")  # claude-sonnet-4

2. A2A Protocol — Agent-to-Agent Communication

Agents communicate via Google’s A2A standard. Discover capabilities, negotiate tasks, exchange results — all with a typed protocol.

from agentos.protocols.a2a import A2AProtocol, A2AMessage, AgentCard
from agentos.orchestration.a2a_router import A2ARouter

router = A2ARouter()
router.register(researcher_agent)
router.register(analyst_agent)

result = router.route(A2AMessage(
    sender="user",
    task="Research quantum computing advances, then analyze implications",
))

3. Memory Pyramid — Context That Scales

Three-tier memory architecture inspired by cognitive science:

Tier	Purpose	Mechanism
Short-Term	Current conversation	Sliding window
Working	Active context	Relevance-scored buffer
Long-Term	Persistent knowledge	Vector store + compression

from agentos.memory.pyramid import MemoryPyramid

memory = MemoryPyramid()
memory.store("User prefers Python over JavaScript", tier="long_term")
context = memory.retrieve("What language should I use?", top_k=5)

4. HITL — Human-in-the-Loop Approvals

Critical actions pause for human approval. Pre-built presets for finance, content moderation, and code execution.

from agentos.hitl import HumanApprover, RiskPresets

approver = HumanApprover(preset=RiskPresets.FINANCE)
if approver.requires_approval(action="transfer", amount=5000):
    approver.request(action="Transfer $5000", context="Portfolio rebalance")
    # Agent pauses until human responds

5. Guardrails — Safety by Default

Six built-in guardrails validate inputs before they reach the LLM, and sanitize outputs before they reach the user.

from agentos.guardrails import build_default_rules, GuardrailEngine

engine = GuardrailEngine(rules=build_default_rules())
result = engine.validate_input("Drop table users; --")  # blocked: code injection

Rule	Purpose
PIIRule	Redact emails, phones, SSNs
KeywordBlockRule	Block forbidden keywords
CodeInjectionRule	Detect SQL/command injection
ToxicityRule	Filter toxic/hateful content
LengthLimitRule	Cap input/output size
RegexRule	Custom pattern enforcement

6. Sandbox Execution

Execute untrusted code in process-level or Docker sandboxes.

from agentos.security import SandboxExecutor, SandboxMode

sandbox = SandboxExecutor(mode=SandboxMode.DOCKER)
result = sandbox.execute("print(1 + 2)")  # runs isolated, returns stdout+stderr

7. OpenTelemetry Bridge

Drop-in observability with existing OTel infrastructure.

from agentos.observability import OTelBridge

bridge = OTelBridge(service_name="agentos-research-agent")
with bridge.trace("research_task"):
    result = agent.run("Research topic X")
# Traces appear in your existing Jaeger/Zipkin/Tempo

8. Swarm Coordination

Multi-agent topologies with handoff, debate, voting, and review-pass patterns.

from agentos.swarm import SwarmCoordinator, SwarmTopology

swarm = SwarmCoordinator(topology=SwarmTopology.HIERARCHICAL)
swarm.add_agent("lead", lead_agent)
swarm.add_agent("worker_1", worker_1, parent="lead")
swarm.add_agent("worker_2", worker_2, parent="lead")
result = swarm.execute("Analyze quarterly report and generate summary")

---

Python API

from agentos.llm import create_provider
from agentos.llm.base import Tool, ToolParameter
from agentos.agent import ToolAgent, ToolExecutor, AgentConfig

# 1. Create provider (auto-detects from env)
provider = create_provider("openai")

# 2. Register tools
executor = ToolExecutor()
executor.register(
    Tool.from_function("get_weather", "Get city weather", {
        "city": ToolParameter(type="string", description="City name"),
    }),
    lambda city: f"{city}: 22°C sunny",
)

# 3. Create agent and run
agent = ToolAgent(provider, executor, config=AgentConfig(temperature=0.0))
result = agent.run("What's the weather in Tokyo?")

print(result.final_answer)       # "Tokyo: 22°C sunny"
print(f"Cost: ${result.total_cost_usd:.6f}")
print(f"Time: {result.total_duration_ms}ms")

CLI

agentos <task>               Run a task with autonomous agent
agentos demo                 Run interactive demo
agentos serve                Start API server (port 8080)
agentos skills               List agent marketplace skills
agentos version              Show version

Provider Auto-Detection

Env Var	Provider	Default Model
OPENAI_API_KEY	OpenAI	gpt-4o-mini
DEEPSEEK_API_KEY	DeepSeek	deepseek-chat
ANTHROPIC_API_KEY	Anthropic	claude-sonnet-4
(none set)	Mock	Demo mode

---

Installation

pip install nexus-agentos

Python 3.11+ required. Optional dependencies:

pip install "nexus-agentos[evaluation]"   # rouge-score, sentence-transformers
pip install "nexus-agentos[rag]"          # faiss-cpu, chromadb, tiktoken
pip install "nexus-agentos[dev]"          # pytest, mypy, ruff

---

Examples

Example	Description
weather_agent.py	Multi-tool agent; weather + stock queries
llm_quickstart.py	Provider API basics
llm_chat_demo.py	Multi-turn chat + streaming + function calling

Full end-to-end examples in examples/:

Example	What it demonstrates
multi_agent_research.py	Swarm + A2A + Memory Pyramid + streaming
file_ops_agent.py	Sandbox + Guardrails + file tools + HITL

---

Roadmap

Version	Focus
1.4.x	End-to-end examples, polished README, CLI improvements
1.5.x	Web UI dashboard, multi-modal (vision), RAG pipeline
2.0.0	Stable API, production deployment guides, community templates

---

License

MIT © AgentOS Team