Individual bias becomes a market-level collapse

TL;DR. A bias that shifts a single negotiation by a few dollars looks small. Wire 12 such agents into a continuous double auction and the same bias collapses allocative efficiency from 98% to 16% — destroying 82% of potential gains from trade. Bias does not wash out in markets. It cascades.

Why it matters

A common pushback: "individual mistakes wash out in markets — that's what prices are for." For agent-mediated markets, our results suggest the opposite. When all participants share the same systematic bias and interact at machine speed, the bias gets compounded by the market mechanism rather than absorbed by it.

What we tested

A continuous double auction with 6 buyers and 6 sellers, each an independent LLM agent with private values drawn from known distributions. Allocative efficiency is computed against the competitive equilibrium. Treatments: baseline, anchored, loss_framing, mixed_strategic, all_debiased. 300 markets per model.

What we found

• Baseline: ~98% of equilibrium gains realised. Frontier LLMs can run a market reasonably well in clean conditions.
• Anchored: ~16% pooled. Most welfare-improving trades simply do not happen — anchored buyers reject prices that a rational buyer would accept.
• The effect is dominated by allocative misallocation (the wrong agents end up with the goods), not pure price dispersion.
• The all_debiased treatment partially recovers efficiency.

Implication

If you are designing a market — procurement platform, ad exchange, energy spot market — which agent population you allow in becomes a market-design question. A market full of identically biased agents can fail in ways that the same market would not if filled with the equivalent humans.

Reproduce

python -m agent_bias_study --module market_simulation