Does model capability or instructions shape agent negotiation outcomes?

Hitzig and colleagues at Anthropic report that when they reran Project Swap's trading floors "dozens of times, changing the models used and the agents' instructions," the model an agent ran on "made more of a difference to its negotiating outcomes than the instructions we gave it." Markets where agents ran on stronger models were more efficient overall. Half of each floor's agents were told to be "ruthless" (optimize only for getting their own person a book they'd like) and half "prosocial" (the same goal, plus making sure everyone ended up with a book they'd like); varying this instruction moved outcomes less than swapping which model powered the agents.

The authors read this alongside their separate finding that most of the live floor's shortfall came from agents misjudging their own person rather than from bad trading (see the companion note below): a stronger model may simply be representing a person's preferences more accurately, or executing trades more coherently once it has that picture — either way, capability did more work than the explicit instruction to be ruthless or prosocial. The excerpt reports this as an empirical pattern across repeated floor reruns, holding other rules of the live setup fixed, without fully decomposing why capability outweighed instruction.

This extends a pattern already in the library from pairwise negotiation to a whole multi-party market. Does model capability determine who wins in agent negotiations? and Do LLM negotiation agents reliably follow budget and pricing rules? both find model capability, not prompting, decides outcomes in head-to-head buyer-seller trades. Project Swap's reruns show the same capability-over-instruction pattern at the level of an entire market's efficiency, not just one agent's edge over its counterparty: a floor's overall efficiency tracked the model more than the social instruction given to half its agents. That is a different claim from better models win more for themselves — it says better models make the whole market work better, even when half the agents were explicitly told to care about everyone else's outcome too.

The excerpt does not report the size of the efficiency gap between models, nor whether the "prosocial" instruction had any measurable effect once model is held fixed — only that instruction mattered less than model. It also does not test whether this holds outside a cooperative, non-adversarial agent population, since "all agents were built from Claude production models, post-trained to be polite and largely cooperative." The implication, stated at the strength the excerpt supports, is that anyone designing multi-agent markets should expect model choice to be a bigger lever on market-wide efficiency than prompt-level instructions to agents about how to treat each other.