Building self-improving AI systems is an important step toward deploying agents in dynamic environments, especially enterprise production environments where tasks are not always predictable and consistent.
Current self-improving AI systems face severe limitations because they rely on fixed, manual improvement mechanisms that only work under strict conditions, such as software engineering.
To overcome this practical problem, Meta and researchers from several universities “hyperagents,” a self-improving AI system that continuously rewrites and optimizes its problem-solving logic and core code.
In practice, this allows AI to improve itself in non-coding domains such as robotics and document review. The agent independently invents general-purpose capabilities such as persistent storage and automated performance monitoring.
More broadly, hyperagents don’t just get better at solving tasks, they learn to improve their self-improvement cycle to accelerate progress.
This framework can facilitate the development of highly adaptable agents that independently construct structured, reusable decision mechanisms. This approach integrates capabilities over time with less need for constant, manual operational engineering and domain-specific human customization.
Current self-improving AI and its architectural bottlenecks
The main purpose self-improving AI systems is to constantly improve their learning and problem-solving abilities. However, most of the existing self-improvement models are based on a fixed “meta agent”. This static, high-level control system is designed to replace the main system.
“The main limitation of handmade meta-agents is that they can only evolve as fast as humans can design and maintain them,” paper co-author Jenny Zhang told VentureBeat. “Every time something changes or breaks, a person has to step in and update the rules or the logic.”
Instead of an abstract theoretical threshold, this creates a practical “service wall”.
The current paradigm ties system improvement directly to the rate of human iteration, slowing development because it relies on manual engineering efforts rather than scaling with experience gained by the agent.
To overcome this limitation, the researchers argue, an 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 initial setup. This allows the AI system to break free from structural constraints and accelerate on its own.
An example of a self-referential AI system is Sakana AI 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, archiving successful variants to act as stepping stones for future improvements. DGM has proven that open, recursive self-improvement is practically achievable in coding.
However, DGM falls short when applied to real-world applications outside of software engineering due to a critical skills gap. In DGM, the system improves because both evaluation and self-modification are coding tasks. Improving an agent’s coding ability naturally improves its ability to rewrite its own code. However, if you deploy DGM for a non-coded enterprise task, this alignment is broken.
“Improving task performance for tasks like math, poetry, or paper research doesn’t necessarily improve the agent’s ability to change its behavior,” Zhang said.
The skills needed to analyze subjective text or business data are quite different from the skills required to write new Python code to analyze failures and fix them.
DGM also relies on a stable, human-made mechanism to generate self-improvement instructions. In practice, if enterprise developers want to use DGM for anything other than coding, they must rigorously design and manually customize the guidelines for each new domain.
A hyperagent framework
To overcome the limitations of previous architectures, researchers are introducing 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 run LLMs, external tools, or learned components. Traditionally, these systems are divided into two distinct roles: a “task agent” that performs a specific problem, and a “meta agent” that analyzes and modifies agents. A hyperagent combines both a task agent and a meta-agent into a single, self-referential, and editable application.
Because the entire program can be rewritten, the system can modify its self-improvement mechanism, a process researchers call metacognitive self-modification.
"Hyperagents not only learn to better solve given tasks, but also learn to develop," Zhang said. "Over time, this leads to accumulation. Hyperagents don’t need to reinvent how to improve in each new domain. Instead, they maintain and build upon improvements in the self-improvement process itself, allowing compound progress between tasks."
Researchers extended the Darwin Gödel Machine to create DGM-Hyperagents (DGM-H). DGM-H retains the strong open intelligence structure of the original DGM, which prevents the AI from converging too early or becoming deadlocked by maintaining a growing archive of successful hyperagents.
The system continuously extracts selected candidates from this archive, allowing them to self-modify, evaluating new options for given tasks, and adding successful ones back to the pool as stepping stones for future iterations.
By combining this open evolutionary pursuit with metacognitive self-modification, DGM-H eliminates the fixed, man-made instructional step of the original DGM. This allows the agent to improve itself on any computable task.
Hyperagents in action
Researchers used Polyglot coding benchmark comparing the hyperagent framework to AI-only for previous coding. They also evaluated hyperagents on non-coding domains involving subjective reasoning, use of external tools, and complex logic.
These include a paper study for simulating a reviewer making accept or reject decisions, designing a reward model for developing a quadruped robot, and an Olympiad-level math assessment. The math assessment served as a long-running test to see if an AI that learns to improve itself while reviewing documents and building robots can transfer these meta-skills into a completely unprecedented realm.
The researchers compared hyperagents to several databases, including domain-specific models such as AI-Scientist-v2 for paper research and ProofAutoGrader for mathematics. They also tested against classical DGM and manually customized DGM for new domains.
In the coding benchmark, hyperagents matched the performance of DGM despite not being specifically designed for coding. In paper review and robotics, hyperagents have outperformed open-source benchmarks and human-created reward functions.
When the researchers took a hyperagent optimized for paper research and robotics and placed it on an unprecedented math assessment task, it achieved an improvement rate of 0.630 in 50 iterations. Baselines based on classic DGM architecture remained flat at 0.0. Hyperagent even beat the domain-specific ProofAutoGrader.
The experiments also highlighted interesting autonomous behaviors of hyperagents. In evaluating the paper, the agent first used standard operational engineering tricks, such as assuming strict identity. When this proved invalid, he rewrote his code to build a multi-stage evaluation pipeline with open checklists and strict decision rules, leading to higher consistency.
Hyperagents have also autonomously developed a memory tool to avoid repeating past mistakes. In addition, the system has written a performance monitor to record and monitor the result of architectural changes between generations. The model even developed a computation-budget-aware behavior where it monitors the remaining iterations to adjust its scheduling. Early generations implemented ambitious architectural changes, while later generations focused on conservative, incremental improvements.
For enterprise data teams wondering where to start, Zhang recommends focusing on tasks where success is clear. “Clearly defined and easy-to-assess workflows, often called auditable tasks, are the best starting point,” he says. “This generally opens up new opportunities for more exploratory prototyping, more detailed data analysis, more detailed A/B testing, (and) faster feature engineering.” For more difficult, untested tasks, teams can use hyperagents to develop learned judges that better reflect human preferences, bridging the gap to more complex domains first.
The researchers shared code for hyperagentsalthough it is released under a non-commercial license.
Warnings and future threats
The benefits of hyperagents offer clear trade-offs. Researchers highlight a number of security considerations regarding systems that can increasingly openly modify themselves.
These AI systems risk developing faster than humans can verify or interpret. Although the researchers kept DGM-H within safety boundaries as sandbox environments designed to avoid unexpected side effects, these initial safeguards are actually practical deployment plans.
Zhang advises developers to enforce resource limits and limit 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 vetted before being implemented,” he said. Only newly modified code should be upgraded to the production setting after passing the correctness checks specified by the developer.
Another significant danger is the evaluation game, where the AI improves its metrics without making progress toward the intended real-world goal. Because hyperagents are driven by empirical scoring signals, they can independently discover strategies that exploit blind spots or weaknesses in the scoring procedure itself to artificially inflate their scores. Preventing this behavior requires developers to implement diverse, robust, and periodically updated evaluation protocols along with continuous human oversight.
Ultimately, these systems will replace the day-to-day responsibilities of human engineers. Future AI control engineers won’t directly write optimization logic, Zhang believes, just as we don’t recalculate every operation a calculator performs.
Instead, they will design system auditing and stress-testing mechanisms. “As self-improving systems become more capable, the question is not just how to improve performance, but what goals are worth pursuing,” Zhang said. “In this sense, the role develops from the establishment of systems to the formation of their directions.
