Quickstart — Hello agent in 60 seconds

Sagewai's first promise is the cheapest one to verify: install the SDK, run one file, get a working agent. This page is the 60-second proof. Everything else on the docs site assumes you've done this once.

What you get from this page: a runnable Python script, a real agent reply, and a clear sense of where to read next based on what you're trying to ship.

If you want the longer "build a portfolio site with Claude Code in a sandbox" tour instead, jump to Getting started — full quickstart. That page exercises the whole stack — Sealed identities, Docker sandboxes, durable workflows. This page is the lighter touch.

Install

pip install sagewai

That's the only required dependency for the hello-world path. No Postgres, no Redis, no admin server. The SDK ships with a no-LLM-key fallback so the example below runs without any secrets.

Run Example 01

The canonical first example is 01_hello_agent.py. Copy this into a file and run it:

import asyncio
from sagewai import UniversalAgent

async def main() -> None:
    agent = UniversalAgent(name="hello", model="ollama/llama3.2:latest")
    reply = await agent.chat("Say hello in one short sentence.")
    print(reply)

asyncio.run(main())

python 01_hello_agent.py

If you have Ollama running with llama3.2 pulled, the reply is real and free. If not, swap the model string for claude-haiku-4-5-20251001 and set ANTHROPIC_API_KEY — same code, different LLM.

What just happened

You imported UniversalAgent, instantiated it with a model identifier, and called .chat(). Sagewai resolved the model identifier through LiteLLM, dispatched the call, and returned the reply as a string. That is the entire SDK surface for a single-turn agent — five lines.

Three things to notice:

1. Same code, any LLM. The model="ollama/..." swap to model="claude-..." to model="openai/..." is one string. Tool calling, memory, workflows, and directives all preserve this property.
2. No vendor account required. Ollama runs locally; the example needs zero paid keys to demonstrate the SDK working end-to-end.
3. No platform required. No admin server, no worker, no Docker. The SDK is a Python library first; the platform layers on top when you need them.

What to read next

Pick the next page by what you're trying to do.

"I want to ship a real feature this quarter"

You're the audience for the Lighthouse section. Six pages, each a real-world problem with a runnable example and a CFO-readable cost story:

• Train your own model — capture answers from Opus, fine-tune a 3B model, deploy it, never pay per-token again
• Moderation and classification — community moderation with three classifiers + an LLM judge, all local
• Memory and retrieval — semantic checkpoint recall and graph-beats-vector for incident response
• Production multitenancy — Sealed credentials, isolated workers, no cross-tenant leak
• Observability and cost — show your CFO where the AI money goes
• Inference deployment — RunPod, Modal, custom endpoints

"I want to learn the SDK first"

Go to Foundation — ten short examples, each a single SDK capability (tools, memory, workflows, MCP, directives, model swap). 60-150 lines each. Read in order.

"I want the full stack tour"

Go to Getting started — the longer quickstart that brings up the admin server, creates a Sealed Identity, and runs Claude Code inside a Docker sandbox to build a portfolio site.

"I want to go deep on one pillar"

Go to Pillars — capability deep-dives for SDK, Autopilot, Fleet, Observatory, and Training Loop. Each pillar links back to its primary lighthouse and foundation examples.