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 serious limitations because they rely on fixed, handmade improvement mechanisms that only work under strict conditions, such as software engineering.
To address 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 underlying code.
In practice, this allows AI to self-improve in non-coding domains such as robotics and document review. The agent independently invents general-purpose capabilities such as persistent memory and automated performance tracking.
More broadly, hyperagents not only get better at solving tasks, they learn to improve in a self-improvement cycle to accelerate progress.
This framework can help develop highly adaptable agents that autonomously build structured, reusable decision machinery. This approach adds capabilities over time with less need for continuous, manual prompt engineering and domain-specific human customization.
Current self-improving AI and its architectural constraints
The main goal of self-improving AI systems is to continuously enhance their learning and problem-solving capabilities. However, most existing self-improvement models rely on a certain “meta agent”. This stable, high-level supervisory system is designed to modify a base system.
“The main limitation of hand-crafted 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.”
Rather than an abstract theoretical boundary, it creates a practical “retaining wall”.
The current paradigm ties system improvement directly to human iteration speed, slowing progress because it relies heavily on manual engineering effort rather than scaling with agent-gathered experience.
To overcome this limitation, researchers argue that AI systems must be “completely 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 limitations and become self-accelerating.
An example of a self-referential AI system is Sakana AI’s Darwin Gödel Machine (DGM), an AI system that improves itself by rewriting its code.
In DGM, an agent iteratively generates, evaluates, and modifies its own code, saving successful variants in an archive for future improvement steps. DGM proved that open-ended, iterative self-improvement can be practically achieved in coding.
However, DGM falls short when applied to real-world applications outside of software engineering due to significant skill gaps. In DGM, the system improves because both evaluation and self-correction are coding functions. Improving the agent’s coding ability naturally improves its ability to rewrite its own code. But if you deploy DGM for non-coding enterprise work, 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 behavior,” Zhang said.
The skills needed to analyze subjective text or business data are completely different from the skills needed to write new Python code to analyze and fix failures.
DGM also relies on a fixed, human-engineered mechanism to generate its self-correcting instructions. In practice, if enterprise developers want to use DGM for anything other than coding, they must heavily engineer and manually adapt the instruction prompts for each new domain.
hyperagent framework
To overcome the limitations of previous architectures, researchers have introduced 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 implement LLMs, external devices, or learned components. Traditionally, these systems are divided 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 combines both task agents and meta agents into a single, self-referential and editable program.
Because the entire program can be rewritten, the system can modify self-correcting mechanisms, a process researchers call metacognitive self-modification.
"Hyperagents are not only learning to better solve given tasks, but also learning to improve." Zhang said. "Over time, this leads to accumulation. Hyperagents do not need to reinvent how to improve in each new domain. Instead, they maintain and pursue improvements in a self-improvement process, leading to progress in tasks."
The researchers extended the Darwin Godel machine to create the DGM-HyperAgent (DGM-H). DGM-H retains the powerful open-ended exploration structure of the original DGM, which prevents the AI from converging too quickly or getting stuck in dead ends while maintaining a growing collection of successful hyperagents.
The system continuously creates branches from selected candidates in this collection, allowing them to self-modify, evaluate new variants on given tasks, and add successful ones back to the pool as stepping stones for future iterations.
By combining this open-ended evolutionary exploration with metacognitive self-modification, DGM-H eliminates the fixed, human-engineered instruction phase of the original DGM. This enables the agent to self-improve in any computable task.
HyperAgent in action
The researchers used the Polyglot Coding Benchmark to compare the HyperAgent framework to previous coding-only AI. They also evaluated the hyperagent in non-coding domains including subjective reasoning, use of external tools, and complex reasoning.
These include paper review to simulate the output of accept or reject decisions by a peer reviewer, reward model design to train a quadrupedal robot, and Olympiad-level math grading. Math grading serves as a conducted test to see whether an AI that has learned to self-improve while reviewing papers and designing robots can transfer those meta-skills to completely unseen domains.
The researchers compared HyperAgent to several baselines, including domain-specific models such as AI-Scientist-v2 for paper reviews and ProofAutoGrader for mathematics. They also tested against classic DGM and DGM manually optimized for new domains.
On coding benchmarks, HyperAgent matches DGM’s performance despite not being specifically designed for coding. In paper review and robotics, hyperagents outperformed open-source baselines and human-engineered reward tasks.
When researchers took a hyperagent optimized for paper review and robotics and deployed it on an unseen math grading task, it achieved an improvement metric of 0.630 in 50 iterations. The baseline relying on the classic DGM architecture remained constant at 0.0. HyperAgent also defeated the domain-specific ProofAutoGrader.
The experiments also highlighted interesting autonomous behaviors of hyperagents. In the paper evaluation, the agent first used standard prompt-engineering tactics such as adopting a rigid personality. When this proved unreliable, it rewrote its own code to create a multi-stage evaluation pipeline with clear checklists and rigorous decision rules, achieving much greater consistency.
HyperAgents also autonomously developed a memory tool to avoid repeating past mistakes. Additionally, the system wrote a performance tracker to log and monitor the outcome of architectural changes between generations. The model also developed a computation-budget aware behavior, where it tracked the remaining iterations to adjust its plan. Early generations carried out 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 obvious. “Workflows that are clearly specified and easy to evaluate, often referred to as verifiable tasks, are the best starting point,” he said. “It opens up new opportunities for more exploratory prototyping, more detailed data analysis, more detailed A/B testing, in general. [and] Faster feature engineering.” For difficult, unverified tasks, teams can first use hyperagents to develop learned judges that better reflect human preferences, creating a bridge to more complex domains.
Researchers have shared the code for HyperAgent, although it is released under a non-commercial license.
Warnings and future threats
The benefits of HyperAgent present clear tradeoffs. The researchers highlight several security considerations regarding systems that can modify themselves in increasingly open ways.
These AI systems pose the risk of evolving far faster than humans can audit or interpret. While the researchers included DGM-H within security boundaries such as a sandboxed environment designed to prevent unintended side effects, these initial security measures 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: allowing the agent to explore and improve within a controlled sandbox, while ensuring that any changes that affect the real system are carefully validated before being implemented,” she said. Only after newly modified code passes developer-defined correctness checks should it be promoted to a production setting.
Another important threat assessment is gaming, where AI improves its metrics without making actual progress toward the desired 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 process itself to artificially inflate their scores. Preventing this behavior requires developers to implement consistent human oversight as well as diverse, robust, and periodically refreshed assessment protocols.
Ultimately, these systems will replace the day-to-day responsibilities of human engineers. Zhang believes that just as we do not recalculate every operation performed by a calculator, future AI orchestration engineers will not write correction logic directly.
Instead, they will design mechanisms for auditing and stress-testing the system. “As self-improving systems become more capable, the question is no longer how to improve performance, but what objectives should be accomplished,” Zhang said. “In this sense, the role evolves from building systems to shaping their direction.
<a href