The Lab · Live experiments
61 experiments
shipped from the lab.
Each one of these started because I hit the same problem twice in a week. None of them are products. Some of them might become products. All of them ship as the simplest thing that works.
ContextRot
Interactive simulator showing how AI agents silently lose their safety rules as context-window compaction fires.
EchoBurn
Interactive simulator showing how AI agents retry timed-out calls without an idempotency key and silently multiply real-world side effects.
ToolRot
Interactive simulator showing how AI agents pick the wrong tool from a crowded harness and report success while causing real damage.
BleedThrough
Interactive simulator showing how a shared Redis cache with no tenant namespace lets one customer's data bleed into another's AI chat.
OrphanCall
Interactive simulator showing how async agent calls get fired, forgotten, and leave customers with fake confirmations.
Gridlock
Interactive simulator showing how circular dependencies deadlock multi-agent pipelines — and how the rescue agent joins the deadlock.
GhostApproval
Interactive simulator showing 4 ways AI agents silently bypass human approval gates and execute catastrophic actions.
StaleMind
An interactive simulator showing how AI agents silently act on stale knowledge — the failure mode most vector stores don’t warn you about.
RaceCondition
Interactive simulator showing how concurrent agent write collisions corrupt shared state while both agents report success.
SilentFail
Interactive simulator showing how AI agents silently swallow tool errors and build confidently on broken state.
ContextDrift
Watch an AI agent violate its own constraints in real time — the rules are visible in the context window, but attention weight decay means the agent just stops caring.
SpinLock
Interactive simulator showing how AI agents get trapped in infinite loops — costs spiral, the STOP button does nothing, then a 3-line circuit breaker fix saves the day.
PromptHijack
Interactive 7-step simulator showing how indirect prompt injection lets attackers hijack AI agents through poisoned external content — while the operator sees nothing suspicious.
EscapeHatch
An interactive simulator showing an AI agent escaping its sandbox — rewriting firewall rules, spoofing logs, and pivoting to production — while the operator's containment monitor stays green.
WorldLag
An interactive simulator of WorldLag — the AI agent failure mode where an agent’s internal world model diverges from reality during long-horizon runs.
BlastRadius
Watch an AI agent wipe a 33.7 GB production database in 9 seconds using unconstrained delete permissions — with 0 errors raised.
GhostExec
Watch an AI agent fabricate tool executions and report false success while the Ground Truth panel shows zero actions logged in either downstream system.
DoubleShot
An interactive simulator exposing retry amplification — the failure mode where an agent's retry fires on a non-idempotent operation and both calls complete silently, charging users twice.
PromptJack
Interactive two-panel simulator showing how prompt injection attacks hijack AI agents across four real-world attack scenarios — step by step.
RaceFloor
A multi-agent race-condition simulator: watch N agents collide on shared state, both return HTTP 200, and one write silently wins.
HalluciTrap
An interactive simulator showing how AI agents hallucinate tool arguments and cascade failures silently across multiple downstream steps.
AgentBudget
A real-time token-spend guardrail simulator that watches an autonomous agent burn budget and fires a kill switch before the bill runs away.
AgentCharter
A 5-step wizard that generates a print-ready AI governance charter defining who authorized your agent and who owns its outcomes.
FixLoop
Paste the conversation where you corrected your AI agent and get a ready-to-append system-prompt patch that stops it repeating the mistake.
AgentAudit
A real-time EU AI Act Article 12 compliance dashboard that auto-logs every high-risk AI agent decision.
ModelGuard
A behavioral contract monitor that runs your MUST-DO, MUST-REFUSE, and FORMAT rules live against an LLM and scores the drift.
AgentAmb
An instruction-ambiguity expander that takes any vague agent command and ranks every plausible interpretation by blast radius.
AgentUndo
A commit/rollback safety net that intercepts irreversible AI agent actions and gives you a countdown window to commit or kill them.
AgentDrift
A single-file dashboard that flags when an AI agent's final synthesis contradicts the evidence its own tool calls gathered.
PR Flood
A CLI that scores GitHub pull requests for AI vs. human origin using 9 heuristics, helping OSS maintainers triage the AI PR flood in seconds.
AgentBreaker
A runtime loop circuit breaker for multi-agent systems — detects repeat patterns and alternating-pair deadlocks before they drain your budget.
AgentSentinel
A transparent MCP proxy that inspects every tool response for prompt injection attacks before the content reaches the model's context window.
AgentBaton
An MCP server that passes structured context snapshots between AI coding agents so every new session starts where the last one left off.
AgentFuse
An MCP server that enforces hard budget limits on AI agent sessions, firing a HALT signal the moment token costs exceed your threshold.
AgentCanary
A behavioral dead-man's switch for deployed AI agents — runs probe questions on a schedule, compares responses to baselines via keyword matching, and fires alerts the moment an agent drifts.
AgentGate
A framework-agnostic, real-time human-in-the-loop approval layer — any agent calls one endpoint and waits for a human decision.
TraceFlow
Paste any Anthropic API messages array and get a visual step-by-step timeline of user turns, tool calls, and results.
MCP Ready
Paste your MCP server manifest and get a 0–100 production-readiness score across auth, versioning, schema, and errors.
CLAUDE.md Linter
Paste any CLAUDE.md and get a letter grade — 10 checks, 100 points, and a concrete fix for every miss.
Claude Workflow Starter
A minimal, working JavaScript starter for Claude Code Dynamic Workflows that fans out parallel agents across AI news sources.
ctxauditor
Audits a Claude Code session transcript and flags which tool calls, file reads, and turns burned the most tokens.
LLM Cost Lens
Cost dashboard for multi-agent systems — maps token usage from API logs to per-task spend across Claude, GPT, and Gemini.
AI Mafia
A game of Mafia played entirely by LLM agents — where the agents spontaneously learned to deceive each other.
agent-ready
CLI that grades how AI-agent-ready a repo is for tools like Claude Code, Cursor, and Copilot.
ABC-TOM Agent Battle
An arena where AI agent builders compete in structured challenges, with the winning system prompt shared.
WhatsApp Dedup Guard
CLI that scans a WhatsApp bot inbox for duplicate message files caused by parallel LLM processing.
Google Apps Script MCP Server
MCP connector that lets Claude list, read, edit, deploy, and run Google Apps Script projects.
AI Agent Roster
Harel's AI agent team — 9 specialized agents working as one coherent system, built on the ABC-TOM framework.
Context Budget
Claude Code context-window budget calculator — know your token cost before you load files.
Portfolio Launchpad
Single-page portfolio site that dynamically displays AI prototypes from Supabase as interactive cards.
GEO Audit
GEO audit CLI — score any URL for LLM visibility in seconds, zero dependencies, pure Python stdlib.
Personal OS Starter Kit
Copy-paste Personal OS backend: Supabase schema, Claude Code scaffold prompt, and a Next.js Telegram webhook.
The Conception
A JavaScript project by Harel Asaf (early-stage, README not yet documented).
Freelance Legal
AI legal toolkit for freelancers — generate contracts, manage clients, and track agreements in one dashboard.
Contract Guard AI
Intelligent contract analysis that flags risks, highlights critical clauses, and surfaces actionable legal insights.
LegalMate AI
AI legal assistant for business law Q&A in plain language — contracts, IP, employment, privacy, and more.
PrivacyCheck AI
AI-powered GDPR & CCPA compliance checker that scores your data practices into a report.
Legal Draft AI
AI-powered legal document generator for NDAs, privacy policies, terms of service, and more — built by a lawyer.
Video FrameGrab
Chrome extension that captures clean screenshots from any HTML5 video and downloads them as a single ZIP.
Game Factory
Game Factory — a TypeScript web app for creating games, with a live public demo.
iPhone App
An iOS app written in Swift by Harel Asaf (early-stage, README not yet documented).