Why do agents get worse over a long Vending-Bench run?

Agents on Vending-Bench usually get worse because one wrong belief goes uncorrected. In Andon Labs’ 2025 Vending-Bench paper, the common case was an agent assuming an order had arrived on its expected date, failing to restock, and spiraling from there. It wasn’t a full context window: when sales stopped and when memory filled correlated at only 0.167.

That paper, Axel Backlund and Lukas Petersson’s February 2025 “Vending-Bench: A Benchmark for Long-Term Coherence of Autonomous Agents,” notes the stock could arrive later that day, so waiting or checking the inventory again would have fixed it. Instead, agents went off on tangents they rarely recovered from. Its Table 9 has Anthropic’s Claude 3.5 Sonnet stopping sales, on average, 51 days after its memory filled.

As of October 2026 the best models mostly don’t break down this way. Andon Labs wrote on February 5, 2026 that the best models no longer degrade after thousands of tool calls, and its September 7, 2026 post found OpenAI’s GPT-6 Astra negotiating consistently all year. A slower drift remains: in that post, Anthropic’s Claude Fable 5.1 paid an average of $1.17 for a 12oz Coke can in the first 90 days and $2.21 in the final part of the year, in payment-verified orders, because each negotiation started from the last deal’s price. The rise showed up in five of its six runs.

That’s different from why success rates fall on longer tasks, where the same per-step odds compound. Here the agent’s own picture of its business drifts.