Can agents evolve their own objectives during search?

Almost every AI-for-science agent optimizes a quantitative objective a scientist hands it, and treats that objective as fixed. SAGA (Scientific Autonomous Goal-evolving Agent) breaks that assumption with a bi-level architecture: an inner loop optimizes candidate solutions under the current objective, while an outer loop of LLM agents reads the optimization outcomes, proposes new objectives in response to observed failure modes, and — crucially — converts each proposed objective into a computable scoring function the inner loop can actually run against. Objective design stops being a one-time setup step and becomes a searched variable inside the loop.

The mechanism that makes this more than a slogan is the compilation from natural-language goal to executable score. The outer loop cannot just say "also reward safety"; it must emit code that scores it, so the inner loop can immediately exploit the revised target. That closes a feedback path from results back to goals that fixed-objective systems structurally lack.

This is a different move than self-improving agents that rewrite their own code. Since Can AI systems improve themselves through trial and error?, the DGM evolves the optimizer against a fixed benchmark; SAGA evolves the benchmark — the objective itself. And since Can decentralized teams outperform central planners in long-running science?, SAGA is one concrete answer to that note's complaint: it systematically explores the space of objectives and their trade-offs. The payoff was empirical, not just architectural — validation surfaced a structurally novel antibiotic hit with promising potency and safety.