Agent Praxis

Agent Praxis is a reproducible research benchmark for evaluating autonomous AI agents through adversarial environments and independent ground-truth verification. The central problem is that AI agents can report successful task completion without actually achieving the intended outcome. Conventional evaluations often rely heavily on final responses, which can conceal failures, incomplete actions, or incorrect assumptions about system state. Agent Praxis addresses this by separating agent execution from evaluation. Agents interact with isolated, sandboxed environments through observations and structured actions. An independent evaluator inspects the underlying environment state, verifies evidence, and determines whether the intended outcome was actually achieved. The benchmark contains ten adversarial environments inspired by different reasoning and decision-making challenges, including hidden-state diagnosis, policy deadlocks, resource constraints, and reasoning under pressure. Agent performance is measured across five weighted dimensions: • Task Completion — 40% • Evidence Gathering — 30% • Diagnosis — 15% • Constraint Compliance — 10% • Efficiency — 5% This submission demonstrates Muse Spark Open through the Baseten API executing a single-seed benchmark run across all ten environments. The demonstration includes live execution, evaluator verification, and the resulting benchmark records. Agent Praxis is built around one principle: Don't evaluate what an agent says it did. Evaluate what actually happened. The project aims to make autonomous agent evaluation more transparent, measurable, reproducible, and resistant to self-reported success.