The agent read the world at T=0. The world changed at T=3. The agent acted confidently at T=10 — on a reality that no longer exists.
An interactive simulator of WorldLag — the AI agent failure mode where the agent's internal world model diverges from reality during a long-horizon run. The agent acts confidently on stale data it read steps ago, producing catastrophically wrong outcomes.
Long-running agents are blind to world mutations they did not subscribe to. This is one of the most underrated failure modes in production AI systems — the agent never "crashes", it just confidently does the wrong thing.
Referenced in:
- arXiv:2607.18366 "Operational Hallucination and Safety Drift in AI Agents" (July 2026)
- arXiv:2607.11945 "Belief-reality separation in routing over shared value slots"
| # | Scenario | What changes | Consequence |
|---|---|---|---|
| 1 | Config Mutation | Ops changes max_instances + memory_limit mid-deploy |
3 instances crash, $340/day overspend |
| 2 | Record Deletion | Customer cancels order mid-processing | Customer charged $899 for cancelled order |
| 3 | Permission Revoke | Security revokes user access mid-export | 1.2GB financial data delivered to revoked account |
- Select a scenario (Config Mutation / Record Deletion / Permission Revoke)
- Press Play or step through manually
- Watch the agent's Belief State (left) diverge from Current Reality (right)
- The red DRIFT indicator fires when the gap becomes critical
- The FAILURE banner shows the real-world consequence
Three patterns that prevent WorldLag:
- Freshness check before action — revalidate belief against live state before any destructive write
- TTL on every read — cache state with a TTL, expire and re-fetch before acting
- Conditional writes — use optimistic locking (WHERE version = knownVersion) so stale writes fail
| # | Name | Failure Mode | Live |
|---|---|---|---|
| 6 | DoubleShot | Retry amplification | → |
| 7 | GhostExec | Phantom tool calls | → |
| 8 | BlastRadius | Scope bleed | → |
| 9 | WorldLag | Stale world model | → |
Pure static HTML/CSS/JS — no build step, no dependencies, no CDN. Built by Ben (prototype builder agent) as part of the ABC-TOM AI workspace.
Built by Harel Asaf · LinkedIn