A Multi-Agent AI Game Cost $1,847 in One Weekend — Here's Why

Horcrux Hunt, a two-agent Harry Potter-themed game built on Claude via Amazon Bedrock, racked up $1,847 in cloud costs over a single weekend demo — more than the developer's rent. Beyond the bill, performance suffered: 12-second latency per turn, a 23% win rate for the protagonist agent, and an 18% timeout rate. Analysis revealed cost isn't additive across agents and turns — it's multiplicative, driven by tokens, agent count, turns, retries, and repeated context replay.

The root causes are instructive for anyone running LLM agents in production: re-sending full conversation history each turn compounds context size exponentially, output tokens cost far more than input tokens and can't be parallelized, and what looks like a single 'turn' actually decomposes into a dozen sequential operations spanning state loading, validation, and persistence. On top of that, 15% of LLM responses failed validation, triggering full-cost retries after the model had already reasoned (and charged) for seconds.

The fix wasn't fewer agents — it was pushing decisions out of the expensive LLM layer wherever possible: a constraint solver that prunes invalid actions before the model ever sees them, a Bayesian belief map that replaces thousands of narrative tokens with a handful of probability values, and heuristic decision trees that skip the LLM call entirely for obvious moves. The guiding principle: fail fast and fail free, catching errors before the meter starts running.

For engineers building multi-agent systems, the takeaway is that cost and latency scale with architectural decisions around context management, validation ordering, and where reasoning gets delegated — not just with model choice or agent count.