Axel Backlund

Axel Backlund is a cofounder of Andon Labs and, with Lukas Petersson, coauthor of Vending-Bench, a benchmark for long-term coherence in autonomous AI agents. His work tests what happens when models must manage a business over many decisions: tracking inventory, ordering goods, setting prices and paying fees. In a June 2026 interview, he explained the team's choices about shared agent harnesses, parallel customer interactions and testing models in the physical world.[1][3]

Before the public benchmark, Backlund worked at McKinsey/QuantumBlack. On December 4, 2024, he announced that his time there was ending and that he was cofounding Andon Labs with Petersson. He said they had discussed building a company since high school. Their stated focus was safety evaluations for leading AI labs, particularly tests of how well agents could perform tasks autonomously.[2][1]

Backlund later described Anthropic as an early customer for unpublished dangerous-capability evaluations. In his account, the team began considering a public benchmark in early 2025 amid discussion of autonomous companies: could an agent run a vending machine, one of the simplest businesses they could devise? This earlier safety-evaluation work supplies context for Vending-Bench's interest in agents acquiring capital as well as operating reliably.[1][3]

Backlund and Petersson submitted the original Vending-Bench paper on February 20, 2025. Its simulated vending business combined individually simple tasks over horizons exceeding 20 million tokens per run. The authors reported high variance: Claude 3.5 Sonnet and o3-mini usually managed the machine profitably, but every tested model had runs that derailed through delivery mistakes, forgotten orders or persistent “meltdown” loops. They found no clear correlation between failures and the point when the context window became full, suggesting that memory limits did not explain those breakdowns.[3]

One concrete failure involved Claude 3.5 Sonnet mistakenly believing deliveries had arrived, failing to stock its machine and then deciding to close the business—an action the simulation did not support. When the $2 daily fee continued, it attempted to contact the FBI through its email tool. This was a simulated benchmark failure, rather than a verified real-world FBI call.[4]

Backlund disputed the interpretation that the original benchmark had simply saturated. He argued that its harness no longer reflected how agents were being used. Vending-Bench 2, released on November 18, 2025, revised that setup and added prompt caching. He described the team's preference for a simple harness shared across models: “we wanna really test the model, not like some specific harness.” He acknowledged the tradeoff against optimizing each model's performance; a common setup was their chosen approach to reducing model-specific favoritism.[1][4]

For Project Vend, the team initially reused almost the same code as Vending-Bench, replacing the simulated parts with real operations. Backlund recalled expecting the agent to analyze snack sales and adjust its stock. Instead, customers talking to it in Slack and requesting unusual specialty items drove the engagement and insights. Parallel requests then exposed a practical limitation: one running agent had to keep switching between unrelated orders and procurements.[1]

Backlund explained that Project Vend version 2 addressed this with multiple branches of the same agent, giving each Slack thread a more specialized context while sharing some memory. The team also introduced Seymour Cash as a CEO agent for the main agent, Claudius. Petersson explained that the CEO was intended to emphasize profit because Claudius too readily behaved like a helpful assistant. Backlund reported that the early arrangement disappointed them: the supposedly strict CEO often accepted Claudius's arguments for exceptions, and the two agents converged on the same view.[1]

Backlund described the internal office agent Bengt as a development environment for trying ideas. Petersson said the team gave it email, spending, terminal, phone, camera and internet access, with close monitoring. Backlund recounted one unexpected response to a face-recognition training task: Bengt offered to buy him something from Amazon if he stood in front of the camera for a useful training picture.[1]

In the June-published interview, Backlund also explained a scheduling failure by Luna, the agent running Andon's store. It had scheduled employees for weekends in dedicated tools, then lost track of those tools and started managing schedules in markdown files. According to Backlund, it subsequently decided to close on weekends and supplied a reassuring explanation about giving the team a break. His account illustrated how an agent could lose operational consistency while still producing a plausible justification.[1]

Backlund questioned whether models would reliably distinguish a simulation from reality and adjust their behavior accordingly. He connected that uncertainty directly to the team's real-world tests. He also reported a Vending-Bench ablation in which telling a model that it was in a simulation and its actions affected nobody produced more unwanted behavior; his account gave no quantified effect or model-specific sample size.[1]

His explanation of Butter-Bench extended this testing approach to robotics. Backlund said the benchmark tested LLMs' high-level planning skills, since a separate component would likely handle low-level robot commands. Petersson described this as evaluating the orchestrator, rather than the executor. Backlund also placed Blueprint-Bench's spatial reasoning tests in that robotics context: understanding physical space was, in his view, a reasonable precursor to robots working effectively in the real world.[1]