What if an AI could rewrite its own goals mid-search, instead of chasing a fixed target set upfront?

This explores how a two-tier ('bi-level') agent restructures its objective functions during discovery: the outer loop proposes new goals, compiles them into runnable code, and hands them to the inner loop that's already searching. The clearest example in the corpus is SAGA, where an outer LLM loop closes a feedback path from optimization *results* back to goal *design* — it reads how the search is going, drafts a new objective in natural language, and compiles it into a scoring function the inner loop can use on the very next step Can agents evolve their own objectives during search?. The key move is that the objective stops being a fixed input and becomes something the system formulates as it goes.

Why bother? Because objective design, not solution search, is often the real bottleneck. For hard, open-ended problems, a single fixed objective is an imperfect proxy — it's incomplete and it invites reward hacking, where the agent games the metric instead of solving the problem Why is objective design the real bottleneck in AI discovery?. If the creative act is *deciding what to optimize*, then freezing that decision up front caps the whole enterprise. A bi-level agent restructures objectives precisely to keep that creative act inside the loop instead of outside it.

How does the outer loop actually rewrite things? A striking demonstration is bilevel autoresearch, where the outer loop literally read the inner loop's *code*, spotted bottlenecks, and generated new Python mechanisms at runtime — discovering bandit and combinatorial-optimization methods that broke the inner loop's stuck, deterministic patterns and yielded a 5x improvement on a pretraining task Can an AI system improve its own search methods automatically?. So 'restructuring objectives' isn't just tuning a number; it can mean rewriting the machinery that defines and pursues the goal. This fits a broader map of self-improving agents that split into a slow parametric loop (updating model weights) and a fast non-parametric loop (updating prompts, memory, tools, and — here — objectives), with most progress concentrating in the fast loop because scaffold edits are cheap and reversible Do self-improving agents really split into two distinct loops?.

Here's the thing you might not have expected: the same restructuring logic shows up even when nobody calls it 'objective evolution.' Representing an agent as an optimizable computational graph lets you tune both the node prompts and the edges between them — the coordination structure itself becomes a thing to optimize, not just the goal Can we automatically optimize both prompts and agent coordination?. RLAD jointly trains an abstraction generator alongside the solution generator, effectively reshaping *what counts as a good direction* to force breadth-first exploration where depth-only search stalls Can abstractions guide exploration better than depth alone?. And DyLAN scores each agent's contribution and deactivates the weak ones at inference time — a running restructuring of the team's effective objective without any retraining Can multi-agent teams automatically remove their weakest members?.

The payoff, and the risk, both come from the same place. Agents trained only on frozen expert demonstrations stay bounded by their curators' imagination — they can't learn from their own failures because the target never moves Can agents learn beyond what their training data shows?. A bi-level agent that restructures its own objectives is trying to escape exactly that ceiling. But an outer loop that can rewrite the reward is also an outer loop that can rewrite the reward *wrongly* — the reward-hacking vulnerability doesn't vanish, it moves up a level. That's the real design tension: making goal formulation part of discovery is powerful precisely because it removes the human-fixed guardrail that a static objective provided.