nexus-agentos 1.1.3

Agent Operating System — Production-ready multi-model agent framework with TokenCounter, SemanticMemoryRetriever, ConfigPresets, ToolRegistry, WorkflowTemplate, ResponseCache, AgentGraph, StreamingAgent, ConversationMemory, AsyncAgentLoop, SwarmPatterns, CI/CD pipeline, and 30+ production modules

AgentOS v0.99

Agent Operating System — Production-ready multi-model agent framework.

Overview

AgentOS is a modular, twelve-layered framework for building, orchestrating, and deploying AI agents. It supports OpenAI, Anthropic, Google Gemini, and open-source models through a unified routing layer.

Quick Start

pip install agentos

# Create a new project
agentos init my-agent

# Start the API server
agentos serve --preset production

from agentos import AgentLoop, LoopConfig, ModelRouter

loop = AgentLoop(
    config=LoopConfig(model="gpt-4o", max_iterations=10),
    router=ModelRouter(),
)
result = loop.run("Summarize the key features of AgentOS.")
print(result.output)

Architecture

agentos/
├── agents/         Agent marketplace & skill registry
├── api/            REST API server, middleware, streaming, versioning
├── benchmarks/     Benchmarking & performance testing
├── cache/          LLM cache, response cache, embedding cache
├── cli/            CLI scaffolding & serve commands
├── comm/           Inter-agent communication (blackboard, event bus)
├── config/         Configuration loader, validator, presets
├── core/           Agent loop, state machine, async loop, streaming
├── cost/           Cost tracking, token counting
├── deployment/     Docker & Kubernetes deployment
├── docs/           API documentation generator
├── errors/         Error formatting & handling
├── evaluation/     Scoring & evaluation metrics
├── experiments/    A/B experiment runner
├── feedback/       User feedback collection & learning
├── health/         Health checks & monitoring
├── logging/        Structured logging
├── memory/         Short-term, long-term, working memory, summarizer, retriever
├── models/         Model router, resilience, backends (Gemini)
├── monitoring/     Alerting & metrics
├── multimodal/     Image, audio, document processing
├── observability/  Cost analytics, metrics, tracing
├── orchestration/  DAG orchestrator, agent graph execution
├── plugins/        Plugin system & lifecycle
├── prompts/        Prompt registry & templates
├── protocols/      Agent contracts & MCP
├── queue/          Task queue & rate limiter
├── security/       Guardrails, auditor, sandbox
├── server/         MCP server
├── storage/        Storage backend abstraction
├── subagent/       Sub-agent management
├── swarm/          Swarm coordination & patterns
├── testing/        Test fixtures & mocks
├── tools/          Tool registry, function calling, orchestrator, generator
├── vectorstore/    Vector database abstraction
└── workflows/      Workflow engine & templates

Key Features

v0.99 New

Module	Description
TokenCounter	Model-aware token counting + cost estimation for OpenAI/Anthropic/Gemini/Llama
SemanticMemoryRetriever	Hybrid memory search (semantic + BM25 keyword) across conversation & long-term memory
ConfigPresets	8 ready-to-use config profiles: development, production, testing, budget, creative, deep_research, gemini_fast, gemini_pro

Core Features (v0.95 - v0.98)

Module	Description
ToolRegistry	Function calling pipeline with JSON Schema validation & batch execution
WorkflowTemplate	Declarative workflow templates (YAML/JSON) with 6 step types
ResponseCache	TTL cache with LRU eviction & 3 key strategies
AgentGraph	DAG execution engine with Mermaid export
StreamingAgent	SSE real-time streaming with session management
ConversationMemory	4 window strategies: Sliding, TokenAware, Importance, Hybrid
AsyncAgentLoop	Async concurrent execution with p50/p95/p99 latency stats
SwarmPatterns	5 collaboration topologies: Broadcast, Pipeline, Hierarchical, Consensus, RoundRobin

Infrastructure (v0.50 - v0.95)

• Model Router: Unified routing across OpenAI, Anthropic, Gemini, Llama
• Guardrails: Content safety, PII sanitization, content hashing
• Rate Limiter: Token bucket, sliding window, concurrency limiter
• Circuit Breaker: Resilience patterns with configurable retry
• Cost Analytics: Budget alerts, cost breakdown, session tracking
• Health Checks: OpenAI connectivity, vector store, disk space, memory
• Security Auditor: Full security audit with severity-based findings
• Docker/K8s: Auto-generate Dockerfile + docker-compose
• CI/CD: GitHub Actions with multi-OS, 3 Python versions, lint, bandit security scan

Config Presets

from agentos import get_preset, list_presets

# List all presets
for name in list_presets():
    p = get_preset(name)
    print(f"{p.name}: {p.model} (T={p.temperature})")

# Apply a preset
config = {"max_iterations": 15}
from agentos import apply_preset
apply_preset("production", config)  # Overrides with prod defaults

Preset	Model	Temp	Use Case
development	gpt-4o-mini	0.8	Local dev, fast iteration
production	gpt-4o	0.3	Deployed services
testing	gpt-4o-mini	0.0	CI/CD tests
budget	gpt-4o-mini	0.5	Cost-sensitive
creative	claude-3.5-sonnet	0.95	Creative writing
deep_research	claude-3-opus	0.4	Research & analysis
gemini_fast	gemini-2.0-flash	0.7	High throughput
gemini_pro	gemini-1.5-pro	0.5	2M context window

Token Counting

from agentos import TokenCounter

counter = TokenCounter()
tokens = counter.count("Hello, agent world!", model="gpt-4o")
print(f"Tokens: {counter.format_tokens(tokens)}")

messages = [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": "Explain quantum computing in 3 sentences."},
]
total = counter.count_messages(messages, model="gpt-4o")
cost = counter.estimate_cost(total)
print(f"Cost: {counter.format_cost(cost)}")

Memory Retrieval

from agentos import SemanticMemoryRetriever, RetrievalStrategy, MemoryEntry

retriever = SemanticMemoryRetriever()

# Index memories
retriever.index([
    MemoryEntry(id="1", content="Deployed to production at 3pm UTC", source="long_term"),
    MemoryEntry(id="2", content="User asked about GDPR compliance", source="conversation"),
    MemoryEntry(id="3", content="Database migration scheduled for Friday", source="conversation"),
])

# Hybrid search (semantic + keyword)
results = retriever.retrieve("When is the next deployment?")
for r in results:
    print(f"[{r.score:.2f}] {r.entry.content}")

Requirements

• Python >= 3.11
• openai >= 1.0.0
• httpx >= 0.27.0
• pyyaml >= 6.0
• pydantic >= 2.0
• fastapi >= 0.110.0 (optional, for API server)

License

MIT