---
id: "concept-multi-level-verification"
type: "concept"
source_timestamps: ["00:16:48"]
tags: ["testing", "quality-assurance"]
related: ["entity-claude-code", "concept-risk-segmentation-permissions"]
definition: "Implementing tests that not only verify the agent's output but also verify that the underlying agentic harness maintains safety guardrails after code updates."
sources: ["s46-anthropic-25b-leak"]
sourceVaultSlug: "s46-anthropic-25b-leak"
originDay: 46
---
# Multi-level Verification

## Definition
Implementing tests at **two distinct levels**: (1) the agent verifying its own outputs, and (2) verification that the agentic harness *itself* still enforces safety guardrails after code changes.

## Level 1 — Agent Output Verification
The expected pattern: the agent has an explicit step to verify its own work. Example: running tests after writing code.

## Level 2 — Harness Verification (the more important one)
When developers make changes to the underlying harness — the plumbing that runs the agent — they need confidence they haven't broken the agent's ability to function safely.

[[entity-claude-code-d46|Claude Code]] includes specific verification tests to ensure the model still respects common guardrails after a harness update. For example, tests verify that **destructive tools still require explicit approval** (see [[concept-risk-segmentation-permissions]]) after refactors.

## Why It Matters
This ensures the infrastructure supporting the AI remains robust and secure as it evolves. The harness is itself code under test — a property easy to forget when the visible artifact is the model's behavior.

## Validation (Enrichment)
Valid. Agent harness tests (e.g., guardrail verification post-update) appear in Pytest suites for frameworks like Semantic Kernel.