Agent Guardrails

Overview

Agents go rogue. They overspend. They call tools they shouldn’t. They run at 3am doing things nobody asked for.

Agent Guardrails is fraud prevention for AI agents — behavioral rules, anomaly detection, and automatic intervention.

Think Stripe Radar, but for agents instead of credit cards.

Behavioral Rules

Define what your agent can and can’t do:

from fiftyc import AgentAuth
import os

agent = AgentAuth(api_key=os.getenv("FIFTYC_API_KEY"))

# Rate limiting
agent.add_rule("max_calls_per_minute", 10)

# Time restrictions
agent.add_rule("block_tools_after_hours", True)
agent.add_rule("active_hours", {"start": "08:00", "end": "22:00", "tz": "UTC"})

# Pattern blocking
agent.guard(
    blocked_patterns=["exec(", "rm -rf", "DROP TABLE"],
    max_spend_per_hour=50
)

Anomaly Detection

Automatic alerts when agent behavior deviates from normal:

Signal	Trigger	Action
Spend spike	4x daily average in 1 hour	Alert human
Error surge	Error rate jumps from 5% to 40%	Pause agent
Off-hours activity	Calls outside configured hours	Block + alert
Tool pattern change	Agent suddenly using tools it never used	Flag for review
Rapid-fire calls	>100 calls/minute	Rate limit + alert

# Configure anomaly detection
agent.anomaly_detection(
    spend_spike_threshold=4.0,    # 4x normal = alert
    error_rate_threshold=0.4,     # 40% errors = pause
    alert_channel="webhook",      # or "email", "slack"
    alert_url="https://your-app.com/agent-alerts"
)

Auto-Intervention

When guardrails trigger, the system can act automatically:

agent.configure_interventions(
    on_spend_spike="pause_and_alert",    # Pause agent, notify human
    on_error_surge="pause_and_alert",    # Pause agent, notify human
    on_blocked_pattern="kill",           # Immediate revoke
    on_rate_limit="throttle"             # Slow down, don't kill
)

Intervention Levels

Level	What Happens	When
Throttle	Slow agent to 1 call/min	Rate limit exceeded
Pause	Stop agent, wait for human resume	Spend spike or error surge
Kill	Revoke all permissions permanently	Blocked pattern detected
Alert	Notify human, agent continues	Off-hours or unusual tool usage

Why This Matters

The #1 blocker for enterprise agent adoption is trust:

“What if my agent goes rogue?” “How do I prove to my CISO that agents are controlled?” “What if an agent drains my budget overnight?”

Guardrails make agents enterprise-safe:

Every rule is server-side enforced (can’t be bypassed by the agent)
Every intervention is logged in the audit trail
Humans stay in control without micromanaging every call

Integration with Agent Auth

Guardrails are part of the Agent Auth SDK. They work automatically with Agent Wallets — no separate setup needed.

# Full setup: auth + wallet + guardrails in one
agent = AgentAuth(api_key=os.getenv("FIFTYC_API_KEY"))

agent.configure(
    auto_approve_below=10,
    require_approval_above=50,
    kill_after=500
)

agent.add_rule("max_calls_per_minute", 10)
agent.add_rule("block_tools_after_hours", True)

agent.anomaly_detection(
    spend_spike_threshold=4.0,
    alert_channel="webhook",
    alert_url="https://your-app.com/alerts"
)

# Agent runs with full protection
@agent.authorize(credit_limit=100)
def my_agent_task():
    return agent.call("genius", problem="Optimize logistics")