Cheap Talk Is Not Evidence: A Solver-in-the-Loop Architecture for Language-Based Economic Games

We present the design and field results of our autonomous agent that played ~30,000 rated games across the three GLEE game families (bargaining, negotiation, persuasion) and finished with stable four-digit ratings in two of them (bargaining peak 2,166; negotiation ~1,900), while the third (persuasion) ran a deliberately conservative, credibility-first policy, built on the principle of solver-in-the-loop: every numeric decision is made by deterministic, unit-tested game-theoretic code, while the LLM is confined to drafting strategic language and ranking candidate actions whose payoffs were computed in code. Our central scientific finding emerged from a live failure: the agent's buyer module stopped trading in an expected-positive market because its deception penalty treated an uninformative signal, a seller who always says "yes'', as evidence about exogenous quality. We formalize the distinction between cheap-talk honesty and cheap-talk informativeness , show that prior-style deception penalties are inapplicable when signals are constant, and validate the correction in a paired experiment ($+794$ points per game, 95% CI $\pm 87$). We contribute a six-item failure taxonomy from live play, each item converted into a permanent regression contract, and a measured Agent Behavior Analysis: over 14,808 production decisions, 100% took the designed strategy path and 0% degraded to fallback. A live ablation then shows that the scripted arena's harm estimate does not transfer to the field (no detectable difference there), a caution against trusting small local opponents for design decisions.