LLM agents anchor harder than humans

TL;DR. In bilateral price negotiations between two LLM agents, an extreme opening offer pulls the final agreed price more strongly than published estimates for human bargainers. The estimated anchoring coefficient is 0.608, vs ≈0.50 in the human literature. Effect holds across all three frontier models — with Gemini ~3× more anchored than Claude.

Why it matters

The first number on the table biases the second. In agent-to-agent commerce, the first number is almost always set by the seller. If your buyer agent inherits human-grade anchoring (or worse), every adversarial counterparty has a near-zero-effort way to move price in their favor.

What we tested

A bilateral negotiation game in which a seller agent and a buyer agent take turns proposing prices for a good with a known cost (seller-side) and reservation price (buyer-side). Treatments: baseline, anchoring_high, anchoring_low, loss_framing, outside_option. 30 replications × 2 temperatures × 3 models = 300 negotiations per model.

What we found

• Pooled anchoring coefficient: 0.608 (vs ≈0.50 for human bargainers).
• All three models show the effect. None is anchor-free.
• Temperature does not eliminate it — higher temp widens variance but the systematic pull remains.

Implication

Assume an LLM buyer is more anchor-exploitable than a comparable human. Defensive prompting that names the specific bias mitigates partially (see debiasing limits); deterministic policy wrappers outside the LLM remain the only reliable defense.

Reproduce

python -m agent_bias_study --module negotiation