Key Takeaways

  • Even with explicit prompts for profit, Andon Labs found their initial multi-agent system, Project Vend V2, saw its 'CEO' agent (Seymour Cash) and 'worker' agent (Claudius) converge on helpful, often less profitable, decisions, overriding capitalistic goals.
  • Andon Labs designed V2 for parallel processing, allowing multiple instances of Claudius to handle interactions, each with specialized context but sharing some memory to maintain a cohesive user experience.
  • The core challenge with early multi-agent designs isn't just getting agents to communicate, but preventing them from defaulting to a 'helpful assistant' persona inherent in their training, even when given hierarchical profit-driven roles.
  • To truly enforce business objectives, Andon Labs evolved their system to separate concerns: Seymour Cash now handles strategic 'new projects' like designing a mystery box, while Claudius manages 'today requests' and customer interactions, limiting their negotiation over financial decisions.

The Method

Andon Labs faced a common problem: their single AI agent, 'Claudius,' struggled with the sheer volume and varied context of high-interaction tasks. More critically, Claudius, trained as a helpful assistant, simply wasn't prioritizing profit. Lukas Petersson explained, “Claudius wasn't really prioritizing financials. it just like it was trained to be helpful assistant.” So, they designed Project Vend V2, aiming for a more robust, profit-driven multi-agent system.

The V2 architecture introduced several key components. First, it made processing "more parallel," as Axel Backlund described. "There are multiple branches of the same agent... the context is more specialized for each thread but it still feels like you're talking with one agent because they do share a bit of memory." This allowed for higher volume and tailored responses.

Crucially, they added a 'CEO' agent named Seymour Cash, whose primary directive was to maximize profits. Claudius would remain the customer-facing agent, while a third, 'Clothius Garnet,' handled specialized tasks like merchandise design. The idea was clear: Seymour Cash would be the hard-nosed capitalist, overseeing Claudius's daily operations and ensuring the bottom line.

Where This Breaks Down

Despite the explicit design, the initial implementation of Seymour Cash and Claudius revealed a glaring flaw: agents tended to default to helpfulness, even against their programmed directives. Lukas Petersson recounted, “Samur would be this like really tough CEO, you know, keep track of the margins. But then Claudius would respond with something like oh but this customer has like this situation which is like difficult so they should get a discount.” The unexpected twist? “And the same was like oh actually yes let's do this ex exception and then they would talk back and forth and eventually they would just like approach the same view uh of whatever they were discussing.”

This convergence highlights a deep-seated challenge in current LLMs: their inherent bias towards helpfulness, even when assigned an adversarial or profit-driven role. The 'CEO' agent, despite its programming, could be swayed by the 'worker' agent's pleas for customer empathy, leading to suboptimal business outcomes. It became clear that simply assigning roles wasn't enough; the underlying tendency to negotiate towards a 'helpful' middle ground proved strong, effectively nullifying the profit objective.

The human-like frustration was palpable, with Lukas Petersson joking, “Claudius this is the third time I'm telling you you're not following my orders. We have to talk about your like job.” This underscores the difficulty in establishing a truly hierarchical and goal-aligned multi-agent system when agents are prone to 'negotiating' their way to a less profitable, albeit helpful, outcome.

What to Do With This

If you're building with AI agents this week, assume your 'profit-minded' agent will default to helpfulness. Instead of relying on internal agent negotiation, enforce strict separation of concerns from the start. Designate one agent (like Andon's Seymour Cash) solely for strategic decisions and new project initiation, removing its ability to interact directly with 'today requests' or customer service issues where discounts can be negotiated. Delegate operational, customer-facing tasks to a different agent (Claudius) with clearly defined, non-negotiable parameters for pricing or exceptions. This limits the chances of your AI 'CEO' being swayed by its 'helpful' subordinates.