Can an AI write its own report card, inventing the goals it's graded on instead of getting them from humans?
Can agents design their own objective functions as part of learning?
This explores whether an agent can write or revise the scoring rules it is optimized against, rather than receiving them from humans, and what that looks like in practice.
This explores whether an agent can write or revise the scoring rules it is optimized against, rather than receiving them from humans. The corpus says yes, in a narrow but concrete form. SAGA uses two nested loops. An inner loop optimizes against a set of goals. An outer LLM loop looks at the results, proposes new objectives in plain language, and compiles them into executable scoring functions the inner loop can use immediately (Can agents evolve their own objectives during search?). Goal design stops being a fixed input and becomes part of the discovery process itself.
A benchmark called Aspire shows why this matters. Agents were given only a vague capability direction, with no tasks and no metrics. They didn't start optimizing. They first spent effort turning the goal into training and validation signals they could measure (Can agents learn from vague goals without predefined metrics?). So objective design is a hidden first phase of learning, which most methods skip because a human already did it.
Several neighboring lines of work loosen the same human-set constraints from different angles. Training on expert demonstrations caps an agent at what its curators imagined (Can agents learn beyond what their training data shows?). One escape is 'early experience'. The agent treats the future states produced by its own actions as supervision, with no reward function at all, and matches expert-dependent baselines with half the data (Can agents learn from their own actions without external rewards?). Reflexion goes a step further. The agent gets only a success or failure signal, then writes its own diagnosis of what went wrong (Can agents learn from failure without updating their weights?). Neither writes a new objective, but both have the agent produce part of its own learning signal. By contrast, RLVMR still relies on human-designed programmatic rewards for planning and reflection (Can RL agents learn to reason better, not just succeed?). That marks the line between rewards an agent is handed and rewards an agent builds.
One survey places this on a longer trajectory. Co-evolving systems shed human engineering in three stages: dynamic peers first, then adaptive environments and feedback, and finally the evolution mechanism itself (Can agents evolve beyond the constraints humans engineer?). Self-written objectives sit in the middle stage, where feedback adapts instead of staying fixed. It also helps that SAGA's objectives are code, not weights. Recent progress in self-improving agents concentrates in the fast loop of prompts, memory, and tools, because those updates are cheap and reversible (Do self-improving agents really split into two distinct loops?). A bad objective can be thrown away.
The retrieved material doesn't show what stops an agent from writing itself an easy objective and declaring victory. What it does suggest is that the loop matters more than the first draft. Across 17 frontier models, persistence in repeated benchmark-and-edit cycles predicted long-horizon success better than initial quality (What predicts success in ultra-long-horizon agent tasks?). A self-designed objective is a hypothesis to be tested and revised, and it only pays off if the agent keeps running that cycle.
Sources 9 notes
SAGA's bi-level architecture closes a feedback loop from optimization results back to goal design by having an outer LLM loop propose new objectives and compile them into code the inner loop can immediately use, enabling objective formulation as part of discovery rather than a fixed input.
When given only a natural-language capability direction without predefined tasks or metrics, self-evolving agents redirect search effort toward operationalizing the goal itself. Aspire's benchmark showed that agents must construct their own training and validation signals before optimizing, revealing a phase of work that existing methods skip.
Agents trained on static expert datasets cannot learn from their own failures or generalize beyond demonstrated scenarios because they never interact with environments during training. Competence is capped by what curators imagined, not by agent capacity.
Research across eight environments shows that agents can use future states from their own actions as supervision without external rewards, matching expert-dependent baselines with half the data and providing superior warm-starts for subsequent RL training.
Reflexion demonstrates that unambiguous environmental feedback (success/failure) enables agents to write useful self-diagnoses and improve across episodes without parameter updates. The binary signal prevents rationalization, and keeping reflections uncompressed preserves their usability.
Show all 9 sources
RLVMR uses structured meta-reasoning tags (planning, exploration, reflection, monitoring) with programmatic rewards to train agentic RL. This reduces repetitive actions by 31% compared to outcome-only methods while maintaining better generalization than supervised fine-tuning alone.
A survey framework organizes co-evolving systems into three stages that progressively remove human engineering: dynamic peers first, then adaptive environments and feedback, finally the evolution mechanism itself. Single-entity self-improvement stalls in static contexts; co-evolution supplies adaptive pressure across multiple components.
A survey framework organizes self-improving agents into two update mechanisms: slow parametric loops updating foundation model weights, and fast non-parametric loops updating prompts, memory, and tools. Recent progress concentrates in the fast loop because scaffold updates are cheaper and reversible than weight updates.
Across 17 frontier models on 36 expert-curated optimization tasks, repeated benchmark-edit-incorporate cycles within a wall-clock budget proved the dominant success predictor. Most models terminated early or burned budget unproductively; Claude Opus 4.6 stood out as persistent.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Self-Improvements in Modern Agentic Systems: A Survey
- The Red Queen Gödel Machine: Co-Evolving Agents and Their Evaluators
- Agent Learning via Early Experience
- Generalized Agent Iteration: One Formal Framework for Iterative Policy Improvement and Recursive Self-Improvement
- RLVMR: Reinforcement Learning with Verifiable Meta-Reasoning Rewards for Robust Long-Horizon Agents
- Dream-RSI: Recursive Self-Improvement through Evolving Worlds
- SkillClaw: Let Skills Evolve Collectively with Agentic Evolver
- Co-Evolution in Agentic Systems: Toward Self-Directed Evolution Beyond Human Design