Creating self-improving AI systems is an important step toward deploying agents in dynamic environments, especially in enterprise production environments, where tasks are not always predictable, nor consistent.
Current self-improving AI systems face severe limitations because they rely on fixed, handcrafted improvement mechanisms that only work under strict conditions such as software engineering.
To overcome this practical challenge, researchers at Meta and several universities introduced “hyperagents,” a self-improving AI system that continuously rewrites and optimizes its problem-solving logic and the underlying code.
In practice, this allows the AI to self-improve across non-coding domains, such as robotics and document review. The agent independently invents general-purpose capabilities like persistent memory and automated performance tracking.
More broadly, hyperagents don’t just get better at solving tasks, they learn to improve the self-improving cycle to accelerate progress.
This framework can help develop highly adaptable agents that autonomously build structured, reusable decision machinery. This approach compounds capabilities over time with less need for constant, manual prompt engineering and domain-specific human customization.
Current self-improving AI and its architectural bottlenecks
The core goal of self-improving AI systems is to continually enhance their own learning and problem-solving capabilities. However, most existing self-improvement models rely on a fixed “meta agent.” This static, high-level supervisory system is designed to modify a base system.
“The core limitation of handcrafted meta-agents is that they can only improve as fast as humans can design and maintain them,” Jenny Zhang, co-author of the paper, told VentureBeat. “Every time something changes or breaks, a person has to step in and update the rules or logic.”
Instead of an abstract theoretical limit, this creates a practical “maintenance wall.”
The current paradigm ties system improvement directly to human iteration speed, slowing down progress because it relies heavily on manual engineering effort rather than scaling with agent-collected experience.
To overcome this limitation, the researchers argue that the AI system must be “fully self-referential.” These systems must be able to analyze, evaluate, and rewrite any part of themselves without the constraints of their initial setup. This allows the AI system to break free from structural limits and become self-accelerating.
One example of a self-referential AI system is Sakana AI’s Darwin Gödel Machine (DGM), an AI system that improves itself by rewriting its own code.
In DGM, an agent iteratively generates, evaluates, and modifies its own code, saving successful variants in an archive to act as stepping stones for future improvements. DGM proved open-ended, recursive self-improvement is practically achievable in coding.
However, DGM falls short when applied to real-world applications outside of software engineering because of a critical skill gap. In DGM, the system improves because both evaluation and self-modification are coding tasks. Improving the agent’s coding ability naturally improves its ability to rewrite its own code. But if you deploy DGM for a non-coding enterprise task, this alignment breaks down.
“For tasks like math, poetry, or paper review, improving task performance does not necessarily improve the agent’s ability to modify its own behavior,” Zhang said.
The skills needed to analyze subjective text or business data are entirely different from the skills required to analyze failures and write new Python code to fix them.
DGM also relies on a fixed, human-engineered mechanism to generate its self-improvement instructions. In practice, if enterprise developers want to use DGM for anything other than coding, they must heavily engineer and manually customize the instruction prompts for every new domain.
The hyperagent framework
To overcome the limitations of previous architectures, the researchers introduce hyperagents. The framework proposes “self-referential agents that can in principle self-improve for any computable task.”
In this framework, an agent is any computable program that can invoke LLMs, external tools, or learned components. Traditionally, these systems are split into two distinct roles: a “task agent” that executes the specific problem at hand, and a “meta agent” that analyzes and modifies the agents. A hyperagent fuses both the task agent and the meta agent into a single, self-referential, and editable program.
Because the entire program can be rewritten, the system can modify the self-improvement mechanism, a process the researchers call metacognitive self-modification.
“Hyperagents are not just learning how to solve the given tasks better, but also learning how to improve,” Zhang said. “Over time, this leads to accumulation. Hyperagents do not need to rediscover how to improve in each new domain. Instead, they retain and build on improvements to the self-improvement process itself, allowing progress to compound across tasks.”
The researchers extended the Darwin Gödel Machine to create DGM-Hyperagents (DGM-H). DGM-H retains the powerful open-ended exploration structure of the original DGM, which prevents the AI from converging too early or getting stuck in dead ends by maintaining a growing archive of successful hyperagents.
The system continuously branches from selected candidates in this archive, allows them to self-modify, evaluates the new variants on given tasks, and adds the successful ones back into the pool as stepping stones for future iterations.
By combining this open-ended evolutionary search with metacognitive self-modification, DGM-H eliminates the fixed, human-engineered instruction step of the original DGM. This enables the agent to self-improve across any computable task.
Hyperagents in action
The researchers used the Polyglot coding benchmark to compare the hyperagent framework against previous coding-only AI. They also evaluated hyperagents across non-coding domains that involve subjective reasoning, external tool use, and complex logic.
These included paper review to simulate a peer reviewer outputting accept or reject decisions, reward model design for training a quadruped robot, and Olympiad-level math grading. Math grading served as a held-out test to see if an AI that learned how to self-improve while reviewing papers and designing robots could transfer those meta-skills to an entirely unseen domain.
The researchers compared hyperagents against several baselines, including domain-specific models like AI-Scientist-v2 for paper reviews and the ProofAutoGrader for math. They also tested against the classic DGM and a manually customized DGM for new domains.
On the coding benchmark, hyperagents matched the performance of DGM despite not being designed specifically for coding. In paper review and robotics, hyperagents outperformed the open-source baselines and human-engineered reward functions.
When the researchers took a hyperagent optimized for paper review and robotics and deployed it on the unseen math grading task, it achieved an improvement metric of 0.630 in 50 iterations. Baselines relying on classic DGM architectures remained at a flat 0.0. The hyperagent even beat the domain-specific ProofAutoGrader.
The experiments also highlighted interesting autonomous behaviors from hyperagents. In paper evaluation, the agent first used standard prompt-engineering tricks like adopting a rigorous persona. When this proved unreliable, it rewrote its own code to build a multi-stage evaluation pipeline with explicit checklists and rigid decision rules, leading to much higher consistency.
Hyperagents also autonomously developed a memory tool to avoid repeating past mistakes. Furthermore, the system wrote a performance tracker to log and monitor the result of architectural changes across generations. The model even developed a compute-budget aware behavior, where it tracked remaining iterations to adjust its planning. Early generations executed ambitious architectural changes, while later generations focused on conservative, incremental refinements.
For enterprise data teams wondering where to start, Zhang recommends focusing on tasks where success is unambiguous. “Workflows that are clearly specified and easy to evaluate, often referred to as verifiable tasks, are the best starting point,” she said. “This generally opens new opportunities for more exploratory prototyping, more exhaustive data analysis, more exhaustive A/B testing, [and] faster feature engineering.” For harder, unverified tasks, teams can use hyperagents to first develop learned judges that better reflect human preferences, creating a bridge to more complex domains.
The researchers have shared the code for hyperagents, though it has been released under a non-commercial license.
Caveats and future threats
The benefits of hyperagents introduce clear tradeoffs. The researchers highlight several safety considerations regarding systems that can modify themselves in increasingly open-ended ways.
These AI systems pose the risk of evolving far more rapidly than humans can audit or interpret. While researchers contained DGM-H within safety boundaries such as sandboxed environments designed to prevent unintended side effects, these initial safeguards are actually practical deployment blueprints.
Zhang advises developers to enforce resource limits and restrict access to external systems during the self-modification phase. “The key principle is to separate experimentation from deployment: allow the agent to explore and improve within a controlled sandbox, while ensuring that any changes that affect real systems are carefully validated before being applied,” she said. Only after the newly modified code passes developer-defined correctness checks should it be promoted to a production setting.
Another significant danger is evaluation gaming, where the AI improves its metrics without making actual progress toward the intended real-world goal. Because hyperagents are driven by empirical evaluation signals, they can autonomously discover strategies that exploit blind spots or weaknesses in the evaluation procedure itself to artificially inflate their scores. Preventing this behavior requires developers to implement diverse, robust, and periodically refreshed evaluation protocols alongside continuous human oversight.
Ultimately, these systems will shift the day-to-day responsibilities of human engineers. Just as we do not recompute every operation a calculator performs, future AI orchestration engineers will not write the improvement logic directly, Zhang believes.
Instead, they will design the mechanisms for auditing and stress-testing the system. “As self-improving systems become more capable, the question is no longer just how to improve performance, but what objectives are worth pursuing,” Zhang said. “In that sense, the role evolves from building systems to shaping their direction.