Yes, in a limited and measurable sense. Recent systems can have an AI model edit the source code of an agent, its tools, or the procedure that improves it, then keep versions that score better on selected tasks. That is different from autonomously redesigning and retraining a smarter foundation model, and it is not yet evidence of a general rise in intelligence.
What “rewriting its own code” means in current research
In the strongest recent demonstrations, a pretrained foundation model remains fixed. It generates changes to an AI agent’s program: the instructions, code-editing tools, context management, evaluation logic, or research workflow around the model. The modified agent is then run on specified tasks. Better-performing variants can be retained and used as the basis for another round.
As an Amazon Associate I earn from qualifying purchases.
The editable layer matters. Changing an agent’s software can make it more effective without changing the neural-network weights that encode the underlying language model. A system that edits training scripts and successfully trains a new foundation model would be a substantially larger claim; the best-known demonstrations below have not established that result.
How the Darwin Gödel Machine works
The Darwin Gödel Machine (DGM) is the clearest example of this approach. Its process is an empirical search rather than a proof that a proposed rewrite must help.
#1 Best Overall
- It selects a coding agent from an archive of existing variants.
- A foundation model proposes a modification to that agent’s code.
- The candidate is run and checked. It must compile and retain the ability to edit a codebase so it can participate in later rounds.
- The candidate is evaluated on coding tasks, and useful variants are added to the archive for possible future selection.
The paper reports changes including improved code-editing tools, better long-context management, and peer-review mechanisms. In its experiments, the system’s score rose from 20.0% to 50.0% on SWE-bench and from 14.2% to 30.7% on Polyglot. Those are the DGM authors’ 2025 benchmark results for coding agents, not a universal intelligence scale.
What DGM did not demonstrate
DGM used frozen pretrained foundation models. Its authors explicitly say that rewriting training scripts to train a new foundation model was not shown because training foundation models is computationally intensive and would add substantial experimental complexity. The demonstrated result is therefore self-improvement of a coding-agent system, not autonomous creation of a more capable general-purpose model.
HyperAgents: making the improvement procedure editable
Meta’s HyperAgents, also called DGM-H, extends the idea by putting a task agent and a meta agent in one editable program. The meta level can modify the procedure that proposes and evaluates improvements, not just the task-solving agent.
Recommended Free Tools
The research page reports experiments in coding, paper review, robotics reward design, and Olympiad-level mathematics-solution grading. It also states that the experiments used sandboxing and human oversight. Those precautions describe the reported research setup; they are not a general guarantee that an unrestricted self-modifying system is safe.
AIDE²: rewriting a research agent’s harness
A September 2026 preprint called AIDE² applies recursive improvement to an AI research agent. Its outer loop rewrites the agent’s harness—the software that organizes the inner loop of task solving—with the goal of making research work more efficient.
The authors report seven accepted successive improvements during an autonomous eight-day run. They also report transfer to four held-out benchmarks, where the resulting system matched or exceeded a human-engineered agent. The paper notes important limits: evaluations can be noisy, and additional runs are costly. These are results from one recent preprint, not a demonstrated law that every agent can improve indefinitely.
How the systems decide whether a rewrite is better
None of these systems can establish in advance that every code change will be beneficial. They generate candidates and use measurements to select among them. The result depends on what is measured, how much evaluation time and compute are available, and which parts of the program are allowed to change.
Archives and selection
DGM keeps an archive of variants rather than relying only on the latest version. That allows a promising branch to be revisited if another branch later performs poorly. Selection is still governed by the tasks and scoring rules chosen by the experimenters.
Rank #3
Benchmark assumptions
DGM treats coding benchmarks as a proxy for coding and self-modification ability. A higher score means the agent solved more of those evaluated tasks under their rules; it does not establish stronger reasoning, knowledge, or adaptability in every setting.
Held-out testing
AIDE²’s transfer to four held-out benchmarks is useful because it tests whether improvements extend beyond the tasks used during the search. Even so, four held-out benchmarks remain a bounded evaluation, not a measurement of general intelligence.
How the approaches compare
| System | What can be edited | Reported evaluation | Selection pattern | What remains unproven |
|---|---|---|---|---|
| Darwin Gödel Machine (2025) | A coding agent’s code, tools, and workflow around a frozen foundation model | SWE-bench and Polyglot coding benchmarks | Archive of variants; candidates are retained after empirical evaluation | Training a new foundation model; general intelligence gain; guaranteed acceleration |
| HyperAgents / DGM-H (Meta, 2026) | The task agent and the meta-level improvement procedure | Coding, paper review, robotics reward design, and Olympiad mathematics grading | Meta-level procedure can itself evolve and is evaluated on selected tasks | Unconstrained self-improvement or a universal intelligence increase |
| AIDE² (September 2026 preprint) | The harness that controls an AI research agent’s inner-loop process | AI research tasks plus four held-out benchmarks | Accepted successive changes during an autonomous eight-day run | Reliable indefinite progress; immunity to noisy evaluations; broad generalization |
Why a benchmark gain is not the same as higher general intelligence
The target of modification is narrower than the headline suggests
An agent can become better at editing files, managing context, or coordinating a research workflow while the underlying model weights remain unchanged. That can be a meaningful engineering improvement, but it is not identical to increasing the model’s general cognitive ability.
Scores reflect the test and the budget
Any measured gain is inseparable from the benchmark design, the scoring metric, the number of candidate rewrites tried, and the compute spent evaluating them. A system optimized for one task family can improve there without becoming better at unrelated tasks.
“Recursive” describes a loop, not an outcome
In this context, recursive means that an improved version can become the next version that proposes or undergoes further changes. The word does not imply exponential progress, uncontrolled behavior, or a guarantee that each generation will be better.
Safety and control in the reported experiments
The cited DGM and HyperAgents work reports controls such as sandboxing and human oversight. These measures limit what an experimental agent can access and provide opportunities for people to inspect or stop runs. They should be understood as safeguards used in those studies, not proof that all future self-modifying systems will remain controllable.
Anthropic’s institutional analysis, “When AI builds itself,” gives a deliberately cautious assessment: We are not there yet, and recursive self-improvement is not inevitable.
It describes possible benefits of more capable systems while also identifying the risk that humans could lose control if full recursive self-improvement were ever achieved.
Quick wins for a faster PC:
Clear out junk files and repair common Windows errorsFree Scan →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →The earlier idea behind today’s agents
A 2022 paper, Self-Programming Artificial Intelligence Using Code-Generating Language Models, described a code-generating model that could modify its own source code and properties such as architecture, computational capacity, and learning dynamics. That work provides historical context for self-programming, but the newer agent-loop studies offer the more relevant evidence for what has been measured recently: iterative changes to software systems evaluated on concrete tasks.
Quick Recap
What the evidence supports
- AI systems can already generate and apply code changes to their own agent or research harness.
- Those changes can produce substantial gains on selected coding, research, or other task benchmarks.
- Some systems can modify the procedure used to search for further improvements.
- Current demonstrations do not show autonomous retraining of a smarter foundation model.
- They do not establish a universal intelligence metric, inevitable recursive acceleration, or a general safety guarantee.
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
