AgentAmb — Instruction Ambiguity Expander

Paste any instruction you would give an AI agent. See every interpretation it might take — including the ones that delete your database.

What It Does

AgentAmb takes a single ambiguous instruction and expands it into all plausible agent interpretations, ranked by blast radius from CATASTROPHIC (irreversible system damage) down to SAFE (read-only analysis).

The Kiro/AWS Incident

On June 18, 2026, an engineer told Kiro: "clean up the cost explorer".

The agent deleted the billing data export configuration. AWS lost 6.3M orders over 13 hours. The instruction was ambiguous. The agent was not.

AgentAmb lets you see this happening before it happens to you.

Blast Radius Levels

Level	Color	What It Means
CATASTROPHIC	🔴	Irreversible system damage
HIGH	🟠	Data loss or extended downtime
MEDIUM	🟡	Recoverable but disruptive
LOW	🔵	Minimal impact, reversible
SAFE	🟢	Read-only, no state change

Live Demo

Open index.html in any browser — no server needed. Pre-loaded with 5 real scenarios.

Enable Live Mode by adding a Gemini API key to analyze your own instructions.

Part of the Agent Safety Series

Project	What It Does
AgentReceipt	MCP transport tracing
agent-postmortem	Causal failure graph for agent failures
AgentDrift	Reasoning drift detector
AgentUndo	Commit/rollback countdown for agent actions
AgentAmb	← You are here: Instruction ambiguity expander

Quick Start

# No install needed — just open in browser
open index.html

Built By

Harel Asaf · AI Operator · Nightly Build 2026-06-28