steer-sdk 0.1.9

The Active Reliability Layer for AI Agents

Steer

steer-labs.com

The Active Reliability Layer for AI Agents.

Stop debugging. Start teaching.
Steer turns runtime hallucinations into permanent fixes—instantly.

Mission Control: Catching hallucinations locally and fixing them with one click.

---

⚡ Quickstart

Get running in 30 seconds. No API keys required.

pip install steer-sdk
steer init   # Generates 3 interactive demo agents
steer ui     # Launches Mission Control dashboard

Then run a demo:

python 01_structure_guard.py

---

🧠 Why Steer?

Most AI tools are passive. Steer is active.

Feature	The Old Way (Observability)	The Steer Way (Reliability)
Reaction	Alerts you after the user crashes.	Blocks the crash before it happens.
Fixing	You edit code, re-prompt, and re-deploy.	You "Teach" the agent a fix in the UI.
Privacy	Sends your prompts to a cloud logger.	Local-First. Data stays on your machine.

---

🛑 The Problem: Logging isn't Enough

When an agent fails in production (e.g., outputs bad JSON or leaks PII), seeing a log doesn't help the user who just got a crash.

Steer creates a "Teaching Layer" around your agent.

1. Catch: Intercept the failure in real-time.
2. Teach: Provide a correction in the Dashboard.
3. Fix: Steer injects that "memory" into the agent immediately.

---

⚡ The Loop: Catch → Teach → Fix

Steer provides a Human-in-the-Loop workflow to fix "Confident Idiot" agents.

1. Catch (The Guard)

Steer wraps your agent and blocks bad outputs before they return.

[Steer] 🤖 Agent generating profile...
[Steer] 🚨 BLOCKED: Structure Guard (Detected Markdown wrapping).
[Steer] 🛡️ Execution halted.

2. Teach (The Fix)

Instead of editing code, go to Steer Mission Control (steer ui).

• Click the blocked incident.
• Click "Teach".
• Select the fix (e.g., "Strict JSON Mode").

Steer now remembers this rule for this agent.

3. Fix (The Result)

Run the agent again. Steer automatically injects your teaching instruction. The agent self-corrects.

[Steer] 🧠 Context loaded: "Strict JSON" rule found. Applying fix...
[Steer] ✅ SUCCESS: Agent output valid JSON.

---

🛡️ Supported Guardrails

Steer comes with 3 "Teach-Ready" verifiers out of the box:

Verifier	The Problem	The "Teaching" Fix
JsonVerifier	Agent wraps code in ```json blocks	"Force JSON": Inject system instruction to output raw JSON only.
RegexVerifier	Agent leaks PII (Emails/Keys)	"Redact PII": Force agent to replace sensitive patterns with [REDACTED].
AmbiguityVerifier	Agent guesses ambiguous answers	"Ask Clarification": Force agent to ask user questions if >3 results found.

---

🛠️ Integration

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. Steer automatically injects rules here!
    # Pass 'steer_rules' to your system prompt.
    system_prompt = f"You are a helpful assistant.\n{steer_rules}"
    
    # ... Your LLM call ...
    return llm.call(system_prompt, user_input)

---

🔮 Roadmap: From Verification to Learning

Steer v0.1 provides the "Fast Path" (Runtime Guardrails + Human Teaching). Steer v0.2+ will introduce the "Slow Path" (Automated Model Improvement).

Coming Soon:

• Query by Committee: Automated consensus checks for ambiguous prompts.
• Automated Fine-Tuning: A pipeline to turn your accumulated (Incident -> Fix) logs into a fine-tuned model that stops making those mistakes entirely.
• CI/CD Integration: Block Pull Requests if an agent fails a reliability test suite.

---

🔑 Configuration (Optional)

The Quickstart demos run locally and require no API keys.

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

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

🧩 Creating Custom Verifiers

Steer is designed to be extensible. You can write your own logic guards in plain Python.

from steer.verifiers import BaseVerifier, VerificationResult, TeachingOption

class MyCustomGuard(BaseVerifier):
    def __init__(self, name: str):
        self.name = name

    def verify(self, inputs: dict, output: str) -> VerificationResult:
        # 1. Your Custom Logic
        # e.g. Check if output contains forbidden words
        is_safe = "competitor_name" not in str(output).lower()
        
        # 2. Return Result
        return VerificationResult(
            verifier_name=self.name,
            passed=is_safe,
            reason="Mentioned competitor name" if not is_safe else None,
            # 3. Define the Fix (that appears in the Dashboard)
            suggested_fixes=[
                TeachingOption(
                    title="Remove Competitor Mentions",
                    description="Detected competitor name.",
                    logic_change="Rule: Never mention competitors. Focus on our product strengths."
                )
            ]
        )