Antaeus

Antaeus is an autonomous research lab for recursive agent improvement. Given a capability goal, Antaeus generates executable synthetic environments that measure the capability, validates those environments with independent auditors, runs agents against them, analyzes where they fail, modifies its own prompts, playbooks, verification rules, or team structure, and promotes a new version only when it beats the previous version on held-out tasks. The central idea is that recursive self-improvement requires a recursive evaluation process: once an agent gets better, its old benchmark stops being informative. Antaeus therefore co-evolves its solver and its environment factory. As capability increases, Antaeus generates harder environments near the new capability frontier. We demonstrate the loop on sandboxed secure-code-repair tasks because they provide rich, composable problems whose outcomes are checked by executing hidden tests. Every agent in the lab runs on Crusoe Managed Inference with NVIDIA Nemotron models: Nemotron 3.5 Lightning is the solver being improved, and Nemotron 3.5 Super builds environments, audits them, analyzes failures, and proposes new agent versions. NVIDIA OpenShell is the execution boundary for all generated code, and a persistent NVIDIA NemoClaw agent acts as the Research Director, steering what the environment factory builds next without ever holding the inference key. The environment factory grounds each defect class in live public context retrieved through the You.com search MCP server.