continuum

by ShyftLabs

The agent runtime for builders who ship.

Build, run, and deploy reliable AI agents at enterprise scale — multi-LLM routing, persistent memory, MCP-native tools, durable workflows, and full observability, out of the box.

📖 Documentation · ⚡ Quick start · ⚙️ Configuration · 🧩 Components · 🧪 Examples · 🤝 Contributing

Continuum is a production-grade Python framework for building, orchestrating, and shipping autonomous AI agents at enterprise scale. It unifies a clean, typed agent core with cost-aware multi-model inference, stateful long- and short-term memory, open standards-based tool calling, durable execution, and end-to-end observability — all behind one small, composable, type-safe API.

✨ Features

• 🤖 Agentic core & orchestration — a strongly-typed agent primitive with full lifecycle hooks, schema-validated structured outputs, and nine composable multi-agent patterns (sequential, parallel, loop, routing, planning, reflection, debate, scatter, supervised).
• 🔀 Smart Inference — cost-aware inference routing that classifies every request by complexity and dispatches it to the cheapest capable model, with seamless cross-provider failover and zero lock-in.
• 🧠 Stateful memory — persistent semantic long-term recall plus low-latency working memory, with multi-tenant isolation scopes and built-in PII redaction for privacy-by-default agents.
• 🔌 Open tool calling — plug into any standards-based tool ecosystem (Model Context Protocol) across multiple transports, with fine-grained capability scoping, context capture/injection, and rich generative-UI artifacts.
• 🔁 Durable execution — long-running, crash- and restart-safe agent workflows with human-in-the-loop approval gates and exactly-once guarantees.
• 🔭 Full observability — first-class distributed tracing, token/latency/error telemetry, and one-line function instrumentation for complete run transparency.
• 🌐 Model-agnostic — target frontier or open-weight models through a single model string; swap providers without touching agent code.
• 🤝 Multi-agent handoffs — context-preserving agent-to-agent delegation with history summarization, cycle detection, and depth control.
• 📡 Real-time streaming — token-, tool-, handoff-, and memory-level events streamed the moment they happen.
• ✅ Built-in evaluation — turn live production traces into golden datasets and regression-test agent quality with standard LLM-evaluation metrics.

🚀 Quick start

Requirements: Python 3.13+ and Docker (for Redis · Milvus/Qdrant · Langfuse).

git clone https://github.com/shyftlabs/continuum.git
cd continuum

python3.13 -m venv .venv && source .venv/bin/activate
pip install -e .

cp .env.template .env        # add your provider key(s) — see Configuration below
docker compose up -d         # Redis · Milvus/Qdrant · Langfuse

Your first agent:

import asyncio
from orchestrator.agent import BaseAgent, AgentRunner

async def main():
    agent = BaseAgent(
        name="hello-agent",
        instructions="You are a friendly assistant.",
        model="gpt-4o-mini",
    )
    runner = AgentRunner()
    response = await runner.run(agent, "Hi!")
    print(response.content)

asyncio.run(main())

AgentRunner.run() returns an AgentResponse with content, structured_output, usage, tool_calls, run_artifacts, latency_ms, and the full handoff chain. See the docs for streaming, tools/MCP, memory, handoffs, and workflows.

⚙️ Configuring Continuum

Continuum is configured through environment variables (copy .env.template → .env). Set keys only for the providers and components you use — everything else has sensible defaults. The most common settings:

LLM providers & routing

Memory (long-term) & embeddings

Sessions (short-term)

Observability (Langfuse)

Temporal (optional, durable workflows)

Optional extras: pip install -e ".[temporal]" for Temporal, ".[eval]" for evaluation, ".[embeddings]" for local embeddings. See .env.template for the complete, annotated reference.

🧩 Components

📚 Documentation

Full documentation lives at docs.continuum.shyftlabs.io — guides for building & running agents, Smart Inference, memory, tools/MCP, workflows, handoffs, streaming, evaluation, and the research behind it.

Markdown sources are also in docs/ if you prefer reading on GitHub — e.g. agent.md, memory.md, tools.md, and the integration GUIDE.md.

🧪 Examples

Runnable demos live under playground/:

• gateway-local-shop — an MCP server + agent + chat UI for a pet-shop assistant (end-to-end: server → agent → UI).
• gateway-multi-agent-shop — a multi-agent workflow variant with routing and handoffs.
• frontend/ — the demo web UIs (assortment, commerce-chat).

🤝 Contributing

Contributions are welcome! Please read CONTRIBUTING.md for the branch model, Conventional Commits, DCO sign-off, and local setup. By participating you agree to our Code of Conduct.

• 🐛 Bugs & features — use the issue templates
• 💬 Questions & ideas — GitHub Discussions
• 🔒 Security — report privately via SECURITY.md, never a public issue

📄 License

Licensed under the Apache License, Version 2.0. Copyright © 2025–2026 ShyftLabs Inc.

For commercial / enterprise inquiries — SLAs, indemnification, hosted offerings, custom features — contact [email protected].