QuantTessera: Evidence-Guided Coding for Quantitative Finance

Quantitative-finance coding agents must produce executable artifacts whose numerical behavior satisfies a task contract. Under bounded model and execution budgets, a successful implementation also depends on how the controller retains evidence and decides when work is complete. We present QuantTessera, a single-model, library-assisted agent that combines heuristic task routing, a deterministic execution-and-repair controller, and task-local memory linked to source versions and tool outcomes. Large observations are stored as retrievable files, while task instructions, active failure summaries, and complete recent action–observation groups are packed into a bounded context. We distinguish three sources of evidence: a public implementation audit, deterministic mechanism tests, and one team-attributed official Agenthon development evaluation. All 29 selected controller tests pass. Across eight synthetic context fixtures, long-log prompt estimates decrease by 79.93–96.26% relative to the implemented legacy policy, whereas short logs incur additional memory and history-retention overhead. The official submission solves 17 of 86 tasks (0.1977); 66 tasks receive a container-failure label and three fail the domain gate. These results support an auditable system design and identify execution reliability as a major observed limitation, without establishing causal accuracy gains from any component.