Steward

Autonomous superagent engine with tool use, multi-provider LLM failover, and deterministic cognitive architecture.

pip install steward-agent[providers]

Quick Start

# Set at least one provider key
export GOOGLE_API_KEY=...        # free tier
export MISTRAL_API_KEY=...       # free tier
export GROQ_API_KEY=...          # free tier
export OPENROUTER_API_KEY=...    # paid (DeepSeek)
export ANTHROPIC_API_KEY=...     # paid (Claude)

# Single task
steward "Fix the bug in main.py"

# Interactive REPL
steward

# Resume previous session
steward --resume "Follow up on the refactor"

# JSON output (machine-readable)
steward --output json "List all TODO comments"

# Telegram bot
pip install steward-agent[providers,telegram]
export TELEGRAM_BOT_TOKEN=...
export TELEGRAM_OWNER_ID=...
steward --telegram

Architecture

Steward follows the Sankhya-25 cognitive model — 24 deterministic infrastructure elements + 1 LLM (the Jiva). The LLM is not the driver; Buddhi (deterministic intellect) controls execution.

User Input
  → Manas (perceive intent, zero LLM)
  → Buddhi (discriminate: tool selection, stuck detection, phase tracking)
  → AgentLoop (LLM call → tool execution → evaluate outcomes)
  → Samskara (context compaction at 50%, LLM summarization at 70%)
  → Response

Key Components

Provider Failover

ProviderChamber manages 5 LLM cells with automatic failover:

1. Google gemini-2.5-flash (free)
2. Mistral mistral-small-latest (free, tool-calling)
3. Groq llama-3.3-70b (free, tool-calling)
4. DeepSeek v3.2 via OpenRouter (paid, $0.27/MTok)
5. Anthropic Claude (paid, highest capability)

Each cell has:

• CircuitBreaker — skip failing providers for 30s after 5 failures
• Cell integrity — membrane degrades on failure, regenerates daily
• FeedbackProtocol — outcome tracking adjusts provider sorting

Safety

• Narasimha killswitch — blocks dangerous shell commands (rm -rf, etc.)
• Iron Dome — blocks blind file writes without prior read
• Buddhi abort — stops stuck loops (3x identical calls, 5x consecutive errors)

Configuration

Optional .steward/config.yaml in your project root:

max_output_tokens: 4096
auto_summarize: true

Environment Variables

Programmatic Usage

import asyncio
from steward import StewardAgent
from steward.provider import build_chamber

chamber = build_chamber()
agent = StewardAgent(provider=chamber)

async def main():
    async for event in agent.run_stream("List all Python files"):
        if event.type == "text_delta":
            print(event.content, end="")

asyncio.run(main())

Part of the Steward Federation

License

MIT