steer-sdk 0.2.8

The Active Reliability Layer for AI Agents

Steer SDK

Active Reliability Layer for AI Agents.

Steer is an open-source Python library that intercepts agent failures (hallucinations, bad JSON, PII leaks) and allows you to inject fixes via a local dashboard without changing your code.

The Problem

When an agent fails in production (e.g., outputs bad JSON), logging the error isn't enough. You usually have to:

1. Dig through logs to find the prompt.
2. Edit your prompt template manually.
3. Redeploy the application.

The Solution

Steer wraps your agent function. When it detects a failure, it blocks the output and logs it to a local dashboard. You click "Teach" to provide a correction (e.g., "Use Strict JSON"), and Steer injects that rule into the agent's context for future runs.

Visual Workflow:

Installation

pip install steer-sdk

Quickstart

Generate the example scripts to see the workflow in action:

steer init
# Generates 01_structure_guard.py, 02_safety_guard.py, etc.

steer ui
# Starts the local dashboard at http://localhost:8000

Run a demo (Split-screen recommended):

1. Run python 01_structure_guard.py. It will fail (Blocked).
2. Go to http://localhost:8000. Click Teach. Select "Strict JSON".
3. Run python 01_structure_guard.py again. It will succeed.

Cookbook

For complex, real-world implementations, explore the cookbook/ directory.

RAG Reliability

This example demonstrates how to enforce strict data schemas and grounding citations in a Retrieval-Augmented Generation (RAG) pipeline.

• Pydantic Schema Enforcement: Ensuring the agent always returns a valid, typed data structure.
• Citation Verification: Hard-locking the agent to cite its sources, preventing ungrounded claims.

View RAG Cookbook

Integration

Steer uses a decorator pattern to wrap your existing functions. To add Steer to your own existing agent, just add steer_rules to your function arguments.

from steer import capture
from steer.verifiers import JsonVerifier

# 1. Define Verifiers
json_check = JsonVerifier(name="Strict JSON")

# 2. Decorate your Agent Function
@capture(verifiers=[json_check])
def my_agent(user_input, steer_rules=""):
    
    # 3. Pass 'steer_rules' to your system prompt.
    # Steer populates this argument automatically based on your teaching.
    system_prompt = f"You are a helpful assistant.\n{steer_rules}"
    
    # ... Your LLM call ...
    return llm.call(system_prompt, user_input)

Data Engine: From Guardrails to Fine-Tuning

Steer does not just catch errors; it creates the dataset needed to fix them permanently.

Every time a rule is applied or an agent succeeds, Steer logs the interaction. You can export these logs into a standard fine-tuning format (JSONL) compatible with OpenAI and other providers.

Export Training Data

Run this command to convert your local logs into a dataset:

steer export

Output: steer_fine_tune.jsonl

Format:

{"messages": [{"role": "system", "content": "..."}, {"role": "user", "content": "..."}, {"role": "assistant", "content": "{\"valid\": \"json\"}"}]}

The Fine-Tuning Workflow

1. Capture: Run your agent with Steer. Fix issues in the Dashboard.
2. Export: Run steer export to generate the dataset.
3. Train: Upload steer_fine_tune.jsonl to OpenAI/Anthropic to fine-tune a model.
4. Remove: Once the model is trained, you can often remove the strict guardrails, reducing latency.

Configuration

The Quickstart demos use a Mock LLM and require no API keys.

To use advanced LLM-based verifiers in production, set your environment variables:

export GEMINI_API_KEY=...
# OR
export OPENAI_API_KEY=...

What is the "Confident Idiot" Problem?

The Confident Idiot is a failure mode where an LLM generates a factually incorrect or structurally broken response with high probability (confidence).

Unlike traditional software that crashes when it fails, LLMs fail silently and plausibly.

• Example: User asks "Weather in Springfield". The agent confidently guesses "Springfield, IL" (ignoring the fact that there are 33 other Springfields in the USA).
• The Fix: Steer prevents this by enforcing Reality Locks (deterministic checks) that run after generation but before the user sees the response.

Read the viral discussion on Hacker News.

Star History