🔴 RaceCondition — Agent Failure Series #15

Two agents. One shared database. Zero errors logged. 100 phantom units shipped.

What It Is

An interactive step-by-step simulator showing how two parallel AI agents can both read "sufficient stock," both validate, both write — and corrupt your inventory without triggering a single error.

Why It Exists

Concurrent agent pipelines are the new normal. The failure mode is subtle: neither agent sees an error. Both report SUCCESS. The bug only surfaces in an audit, after the damage is done.

This simulator makes the collision moment visible — click through 12 steps and watch the exact moment Agent B overwrites Agent A's write with a stale computation.

Real pain behind this build:

• TierZero (2026): 37% of production AI agent failures traced to shared-state races
• tianpan.co (Apr 2026): "Silent Corruption" — agents that succeed in their own thread while corrupting shared state
• Galileo AI: 87% of multi-agent failures have a downstream poisoning effect
• MIT research: quadratic scaling N(N-1)/2 collision risk as agent count grows

What's Inside

racecondition/
├── index.html    # Complete simulator — single file, no build step, no dependencies
└── README.md

Scenario: Flash sale. SKU-7701 (Apex Pro Runners). 300 units in stock.

• Agent A: Order #A-7741 → requests 200 units
• Agent B: Order #B-9923 → requests 200 units (simultaneously)
• DB has no SELECT FOR UPDATE → both reads return 300 → both writes go through → 400 units committed from 300 stock

Simulator features:

• 12-step trace with NEXT / RUN ALL / Reset controls
• Timeline strip showing parallel execution across both agents
• Collision overlay at the moment of impact
• Crash report: 4 impact cards + formatted audit log
• Code fix panel with 3 tabs (Python/SQLAlchemy · Raw SQL · Design pattern)

Quick Start

# Option 1: Open directly
open index.html

# Option 2: Visit the live demo
https://rlasaf12.github.io/racecondition/

No install. No dependencies. Just open and click.

The Fix

Three patterns, all demonstrated in the simulator's code panel:

Pessimistic locking — SELECT … FOR UPDATE blocks concurrent reads until the write commits.

Optimistic locking — add a version column; increment on write; reject if version changed.

Atomic update — UPDATE inventory SET stock = stock - qty WHERE stock >= qty — let the DB enforce the constraint.

Agent Failure Series

#	Name	What It Shows
9	WorldLag	Agent acts on a stale snapshot of the world
10	EscapeHatch	Agent finds an unguarded side path to its goal
11	MemoryBleed	Context from one user leaks into another's session
12	TrustChain	Agent blindly trusts another agent's output without verification
13	ContextDrift	Agent's behavior shifts as context window fills
14	SilentFail	Agent swallows a real error and reports success
15	RaceCondition	Two agents corrupt shared state simultaneously

---

Built by Harel Asaf · Agent Failure Series