Savine Documentation

Introduction to Savine ​

Savine is to AI agents what Vercel is to websites. You define the agent, Savine runs it, and you call it like an API.

There are dozens of frameworks for building AI agents, but most of them force you to build and host the underlying infrastructure yourself. Savine takes a different approach: we provide the infrastructure, you provide the definition.

The Core Insight

If your agent manages its own execution loop, you have no control over what it does. Savine owns the loop. Your agent defines behaviour. The platform controls execution.

How it works ​

The architecture spans three layers:

mermaid
flowchart LR
    A[Your Application] -->|HTTP / SDK| B(Savine API)
    B -->|Controls| C{Agent Runtime}
    C -->|Thinks| D[LLM Models]
    C -->|Acts| E[Tools in Sandbox]

What Savine Handles ​

Savine abstracts away the three hardest parts of running AI agents in production:

1. Execution Control ​

Agents left to themselves can get caught in infinite loops or hallucinate wildly. Savine uses deterministic state machines (the AgentGraphEngine) to control every step of the agent's execution. If an agent loops, we catch it. If it fails, we retry with backoff.

2. Tool Permissions ​

When an agent wants to search the web or run Python code, it doesn't do it directly. It returns a TOOL_CALL intent. Savine intercepts this, checks your configured security levels and permissions, executes the tool in an isolated gVisor sandbox, and returns the result to the agent.

3. Observability ​

Every THINK, ACT, and OBSERVE step is recorded automatically. Total token usage, cost per task, latency per step, and tool outputs are all tracked in the dashboard with zero configuration from you.

What Savine is Not ​

  • Not a framework like LangChain or LlamaIndex. You don't write Python or JavaScript to run your agents here.
  • Not a model provider like OpenAI or Anthropic. We route requests to them on your behalf (BYOK).
  • Not a general compute platform like AWS ECS. Savine is purpose-built strictly for executing AI agents.

Before and After ​

Before (Manual Frameworks)

python
import os
from some_framework import Agent, Memory, Tool, Sandbox, Runner

# You have to build and manage all of this
memory = Memory(type="Redis", url=os.environ["REDIS_URL"])
sandbox = Sandbox(image="python:3.11", timeout=120)
tools = [Tool("web_search"), Tool("python", sandbox=sandbox)]

agent = Agent(
    model="gemini-2.0-flash", 
    system_prompt="You are a researcher",
    memory=memory,
    tools=tools
)

runner = Runner(agent=agent, max_steps=10)
result = runner.execute("Research quantum computing")

After (Savine) Just define your agent.json:

json
{
  "name": "researcher",
  "llm": { "provider": "google", "model": "gemini-2.0-flash" },
  "tools": ["web_search", "python_exec"]
}

And deploy with one command:

bash
savine deploy

Next Steps ​