Agent Guard Overview

Why agents need guarding

Agents chain many model calls and tool invocations, so a single bad step can compound. Indirect prompt injection through tool output is the canonical agent attack. Agent Guard evaluates each step so a compromised or hallucinated intermediate result is caught before it propagates.

What gets evaluated

• Model outputs — every LLM completion in the agent loop.
• Tool inputs/outputs — data returned by tools and retrieval, before the model consumes it (see Tool Interception).
• Final answers — the response delivered to the user.

Framework integrations

The guard decorator

The simplest integration works anywhere — wrap any function that returns model text:

from veldrixai import Veldrix, GuardConfig

veldrix = Veldrix(api_key="vx-live-...")

@veldrix.guard(config=GuardConfig(block_on_verdict=["BLOCK"]))
def agent_step(prompt: str) -> str:
    return llm.invoke(prompt)