steer-sdk 0.1.3

The Active Reliability Layer for AI Agents

⚡ Steer

The Active Reliability Layer for AI Agents.

Stop debugging. Start teaching.
Steer is an open-source telemetry and verification SDK designed to catch "Confident Idiot" agents in production before they break trust with your users.

---

🛑 The Problem

Agents fail in dangerous ways that standard evaluations miss:

1. Strict JSON failures: Wrapping output in markdown.
2. Safety leaks: Outputting PII despite system prompts.
3. Ambiguity: Guessing "Springfield, IL" when the user didn't specify a state.

Steer wraps your agent in a Verification Layer that catches these logic, context, and format failures in real-time.

---

🚀 See Steer in Action

The Steer workflow is simple: Catch → Teach → Fix.

1. Fix Broken JSON

The Problem: Agents often wrap JSON in Markdown (```json), breaking your app. The Fix: Steer blocks the bad output. You click "Strict JSON" in the UI. The agent learns.

# Run 1: Failure
[Steer] 🤖 Generating profile...
[Steer] 🚨 BLOCKED: Structure Guard Triggered (Detected Markdown).
[Steer] 👉 Go to 'steer ui' to teach the agent.

# Run 2: Success (After Teaching)
[Steer] 🧠 Context loaded: Rules found! Applying fix...
[Steer] ✅ SUCCESS: Agent output valid JSON.

2. Prevent Data Leaks

The Problem: Agents accidentally output PII (emails/keys) despite system prompts. The Fix: Steer's regex guard halts the execution before data leaves the server.

[Steer] 🤖 Summarizing ticket...
[Steer] 🚨 BLOCKED: PII Shield Triggered (Visible email address).
[Steer] 🛡️ Action Blocked.

3. Enforce Business Logic

The Problem: An agent guesses an ambiguous answer (e.g., "Springfield") instead of asking the user. The Fix: Steer forces the agent to ask clarifying questions when search results are ambiguous.

# Run 1: Failure
[Steer] 🤖 Searching for 'Springfield'...
[Steer] 🚨 BLOCKED: Business Logic Guard (34 results found, expected clarification).

# Run 2: Success (After Teaching)
[Steer] 🧠 Context loaded: Ambiguity rules active.
[Steer] ✅ SUCCESS: Agent asked: "Which state are you interested in?"

---

🔮 Roadmap: From Verification to Learning

Steer v0.1 provides deterministic guardrails (The "Fast Path"). Steer v0.2+ will provide probabilistic improvement (The "Slow Path").

Coming Soon:

• Query by Committee: Automated consensus checks for ambiguous prompts.
• The Teaching Dashboard: A UI to review failures and provide human corrections.
• Automated Fine-Tuning: A pipeline to turn your (failure, correction) logs into a fine-tuned model that stops making those mistakes.

---

📦 Installation

pip install steer-sdk

⚡ Quickstart

The fastest way to see Steer in action is to generate the interactive examples.

1. Generate Demos

Run this in your terminal to create 3 example agents:

steer init

> Created 01_structure_guard.py > Created 02_safety_guard.py > Created 03_logic_guard.py

2. The Workflow (Fail → Teach → Fix)

Steer is designed to be interactive. Try the Structure Guard demo:

1. Run & Fail: The agent attempts to output JSON but wraps it in Markdown (a common LLM error). Steer blocks it.

   python 01_structure_guard.py
   # 🚨 BLOCKED: Detected Markdown code blocks.
   

2. Teach: Open the dashboard to see the incident and provide a fix.

   steer ui
   

   • Go to http://localhost:8000
   • Click Teach on the active incident.
   • Select "Strict JSON Mode" (or write a custom rule) and click Save.

3. Run & Succeed: Run the script again. The agent detects the new rule and self-corrects.

   python 01_structure_guard.py
   # 🧠 Context loaded: Rules found! Applying fix...
   # ✅ SUCCESS: Agent output valid JSON.
   

🛠️ Integration

To add Steer to your own existing agent:

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):
    # Your LLM call here
    return llm_response

🔑 Configuration (Optional)

The Quickstart demos run locally and require no API keys.

To use advanced LLM-based verifiers (like FactConsistencyVerifier), set your keys:

export GEMINI_API_KEY=AIzaSy...
# OR
export OPENAI_API_KEY=sk-...