🛡️ agent-loop-guard-js

Stop runaway AI agents before they burn your tokens, time, and money.

agent-loop-guard-js is a high-performance, provider-independent safety utility for AI agents. It detects and blocks runaway execution—such as infinite loops, repeated tool calls, and circular patterns—before they happen.

---

🚀 Why do you need this?

AI agents are essentially loops: LLM → tool → LLM → tool. When these loops fail, they fail expensively:

• The Hammer: Calling the same tool with the same arguments 100 times.
• The Error Loop: Hammering a tool that consistently returns an error.
• The Orbit: Falling into circular patterns (Tool A $\rightarrow$ Tool B $\rightarrow$ Tool A).
• The Runaway: Simply never stopping, draining your API credits.

agent-loop-guard-js sits between your agent and its tools, returning a structured allowed: boolean decision. Zero provider dependencies. Zero runtime dependencies. Total control.

---

📦 Installation

npm install agent-loop-guard-js
# or
pnpm add agent-loop-guard-js
# or
yarn add agent-loop-guard-js

---

⚡ Quick Start

import { AgentLoopGuard } from "agent-loop-guard-js";

const guard = new AgentLoopGuard({
  maxSteps: 20,           // Stop after 20 total steps
  maxDuration: 60_000,    // Stop after 60 seconds
  maxRepeatedCalls: 3,    // Stop if same tool + args called 3x
  maxSameToolCalls: 5,    // Stop if any tool is called 5x in a row
});

function beforeEachStep(step) {
  const decision = guard.check(step);
  if (!decision.allowed) {
    throw new Error(`Agent Guard Blocked: ${decision.reason}`);
  }
  // Execute the approved step...
}

---

✨ Features

Protection	Config Key	Detects
Step Budget	maxSteps	Total steps exceeded in a single run.
Time Budget	maxDuration	Run exceeds wall-clock limit (monotonic).
Repetition	maxRepeatedCalls	Identical tool + arguments called repeatedly.
Tool Hammering	maxSameToolCalls	Single tool called consecutively regardless of args.
Circular Loops	loopPatternWindow	Repeating sequences (e.g., A → B → A → B).

Technical Highlights:

• Deterministic Canonicalization: Arguments are normalized so {a:1,b:2} equals {b:2,a:1}.
• Zero-Dependency: No bloat. Works in Node.js, Bun, and Deno.
• O(1) Memory Growth: History is bounded; it will never crash your process on long runs.
• Strongly Typed: Full TypeScript support for a seamless DX.

---

🛠️ API Reference

new AgentLoopGuard(options?)

Constructs a guard. All options are optional (0 = unlimited).

interface AgentLoopGuardOptions {
  maxSteps?: number;
  maxDuration?: number;
  maxRepeatedCalls?: number;
  maxSameToolCalls?: number;
  loopPatternWindow?: number;
  detectDuplicateArguments?: boolean;
  onViolation?: (event: ViolationEvent) => void;
}

Core Lifecycle

• guard.start(): Records the start time of a run.
• guard.check(step): Validates the next step. Returns a GuardDecision.
• guard.end(): Closes the current run.
• guard.reset(): Clears all history for a fresh start.

The GuardDecision

check() returns a decision object:

• allowed: true $\rightarrow$ Proceed with execution.
• allowed: false $\rightarrow$ Stop the agent. Includes a reason (e.g., LOOP_PATTERN_DETECTED).

---

📐 Architecture

flowchart LR
  A[Agent] -->|step| G[AgentLoopGuard]
  G --> D1[MaxSteps]
  G --> D2[Duration]
  G --> D3[Repetition]
  G --> D4[SameTool]
  G --> D5[Circular]
  D1 & D2 & D3 & D4 & D5 --> R{Decision}
  R -->|Allowed| O[Execute Tool]
  R -->|Blocked| V[Stop & Alert]

Loading

---

🤝 Contributing & Support

Contributions are welcome! Please ensure the pipeline is green: npm run typecheck && npm run lint && npm test

If this tool saves you tokens and time, consider supporting its development: 👉 Sponsor on GitHub

📄 License

MIT © OMD-123