Core Concepts

Core Concepts

Understanding AgentWall's key features and how they protect your AI agents.

Run-Level Tracking

What is a Run?

A run is a complete agent task, which may involve multiple LLM requests.

Why It Matters

  • Budget per task: Limit spending per agent task, not just per API key
  • Loop detection: Detect when an agent is stuck across multiple requests
  • Debugging: See the full journey of an agent task

How to Use

Pass agentwall_run_id in your requests:

Loop Detection

The Problem

AI agents can get stuck in infinite loops:

This can cost thousands of dollars before anyone notices.

How AgentWall Detects Loops

AgentWall detects three types of loops:

Type Description Detection
Exact Same prompt repeated 2nd occurrence
Similar Nearly identical prompts Jaccard similarity >85%
Oscillation A→B→A→B pattern 4th occurrence

What Happens When a Loop is Detected

  1. Request is blocked (HTTP 429)
  2. Run is killed (no more requests allowed)
  3. Alert sent to Slack (if configured)
  4. Logged in dashboard

Example Response

Budget Enforcement

Three Levels of Budget

Level Scope Use Case
Per-Run Single agent task "This task can't cost more than $1"
Daily All requests today "Don't spend more than $100/day"
Monthly All requests this month "Monthly budget is $1000"

How It Works

Budget Exceeded Response

DLP (Data Loss Prevention)

What is DLP?

DLP scans requests and responses for sensitive data and masks it automatically.

Detected Patterns

Category Examples
API Keys OpenAI, AWS, Stripe, GitHub, Slack
Credentials Passwords, tokens, secrets
Financial Credit cards (Visa, MC, Amex), IBANs
PII Emails, phone numbers, SSNs
Infrastructure Private IPs, internal URLs

How It Works

Configuration

DLP is enabled by default. You can configure sensitivity in the dashboard:

  • Strict: Block requests with sensitive data
  • Redact: Mask sensitive data and continue (default)
  • Log Only: Log but don't modify

Kill Switch

Manual Kill

Stop a runaway agent immediately from the dashboard:

  1. Go to Agent Runs
  2. Find the run
  3. Click Kill Run

All future requests with that run_id will be blocked.

Automatic Kill

AgentWall automatically kills runs when:

  • Loop detected
  • Budget exceeded
  • Step limit reached (default: 30 steps)
  • Timeout exceeded (default: 120 seconds)

Cost Tracking

Per-Request Cost

Every response includes cost information:

Supported Models

Model Input (per 1K) Output (per 1K)
gpt-4 $0.03 $0.06
gpt-4-turbo $0.01 $0.03
gpt-3.5-turbo $0.0005 $0.0015
claude-3-opus $0.015 $0.075
claude-3-sonnet $0.003 $0.015

Architecture

Components

  • FastAPI Proxy: Ultra-low latency (<10ms overhead)
  • Redis: Run state and loop detection
  • ClickHouse: High-performance logging
  • Laravel Dashboard: Admin panel and alerts

Next: API Reference