Something important is happening in artificial intelligence right now, and I don’t think we should look at it simply as another improvement in AI coding.
AI is increasingly becoming part of the process used to build better AI.
Anthropic recently reported that Claude was leading about 26% of its AI research and development work, up from 1% in March, while more than 90% of its research involved human-AI collaboration in August. Anthropic also says more than 80% of the code merged into its codebase was authored by Claude as of May 2026.
OpenAI has separately reported that its researchers are increasingly using coding agents for research, experimentation, infrastructure and analysis. By mid-August, its research organization was using the equivalent of 3.1 agent-workdays for every human workday.
These numbers are remarkable.
But the number itself isn’t what interests me most.
The question I keep thinking about is:
What happens when AI stops being only the product we are building and becomes part of the team building the next product?
I think that question will become much more important over the next few years.
We Are Moving From AI-Assisted Development to AI-Assisted AI Development
There is an important distinction here.
For the last few years, we have primarily used AI to help us do existing work:
- Write code
- Analyze data
- Draft documents
- Research topics
- Generate tests
- Debug software
- Summarize information
Now we are seeing something different.
AI is increasingly being used to help with the development of AI itself.
That can include:
- Writing research infrastructure
- Running experiments
- Analyzing results
- Generating hypotheses
- Developing evaluations
- Testing systems
- Debugging training pipelines
- Exploring research questions
- Improving software used to develop models
Anthropic describes this as part of a broader progression in which AI systems are moving from executing tasks specified by humans toward taking on increasingly complex goals.
OpenAI is reporting a similar shift in its own research organization, where agents are being used across different stages of the AI R&D lifecycle, including deciding, designing, building, running, analyzing and communicating research.
This is not yet a world where AI independently builds a completely new generation of itself.
We should be precise about that.
Recursive self-improvement in the strongest sense has not arrived.
Anthropic itself explicitly says it is not there yet and that recursive self-improvement is not inevitable.
But we don’t need to reach full autonomy before the development process changes fundamentally. The transition has already started.
The Interesting Part Isn't AI Writing Code
AI writing code is no longer particularly surprising.
The more interesting development is that the role of AI in the development process is expanding upward.
There is a difference between asking:
“Fix this function.”
and:
“Investigate why this system behaves this way.”
And another difference between that and:
“Design an experiment to test this hypothesis.”
And another:
“What should we investigate next?”
The further we move up this ladder, the more interesting—and difficult—the human-AI relationship becomes.
Because eventually the question isn’t simply:
Can AI execute the task?
It becomes:
Who decides what task should be executed?
That is a much bigger question.
The Bottleneck Will Move
This is where my process-excellence thinking comes into the picture.
Whenever we make one part of a system dramatically faster, the bottleneck usually moves somewhere else.
AI can already accelerate coding.
So coding may become less of the bottleneck.
AI can accelerate experimentation.
So running experiments may become less of a bottleneck.
AI can accelerate analysis.
So analysis may become less of a bottleneck.
What happens then?
The bottleneck could move toward:
- Problem selection
- Research direction
- Evaluation
- Compute
- Data
- Safety verification
- Human judgment
- Infrastructure
- Integration
OpenAI itself makes a similar observation in its recent research-acceleration report: as automation progresses, tasks that are less automatable can become a larger share of researcher effort and therefore become important bottlenecks.
This is something I think businesses should pay much more attention to.
AI doesn’t eliminate bottlenecks.
It often moves them.
And That Changes What an AI Researcher Does
Imagine an AI researcher five years ago.
A significant amount of time might have gone into:
- Writing code
- Running experiments
- Debugging infrastructure
- Processing results
- Preparing analysis
Now imagine that AI systems perform a large part of that work.
What happens to the human?
I don’t think the answer is simply:
“The human becomes unnecessary.”
The more interesting possibility is that the human moves upward.
The human spends more time deciding:
- Which problem matters?
- Which experiment is worth running?
- Which result is meaningful?
- Which hypothesis deserves investment?
- Which failure teaches us something?
- Which system should be trusted?
- Which direction should the organization pursue?
In other words:
The value of deciding what to do may increase as the cost of doing it decreases.
That is a major shift.
The Next Scarce Resource May Be Good Questions
We have spent years talking about the scarcity of computing power.
Then we talked about the scarcity of data.
We are now increasingly talking about the cost of inference.
But there is another resource that becomes more important when AI can perform enormous amounts of work:
Good questions.
If an AI system can run thousands of experiments, the difficult question may no longer be:
“Can we run the experiment?”
It may be:
“Which experiment deserves to be run?”
If AI can generate thousands of possible product ideas, the bottleneck becomes choosing the right one.
If AI can write millions of lines of code, the bottleneck becomes deciding what software should exist.
If AI can produce thousands of research hypotheses, the bottleneck becomes determining which hypotheses matter.
This is why I don’t believe human judgment becomes irrelevant as AI becomes more capable.
In some areas, judgment becomes more valuable because AI makes execution abundant.
But There Is a Serious Problem: Who Tests the Tester?
This is where things get complicated.
Suppose AI helps create a new AI system.
That system is then evaluated.
But the evaluation itself may have been created or optimized with AI.
The testing infrastructure may have been written by AI.
The experiment may have been designed with AI.
The analysis may have been performed with AI.
Now we have a system where increasingly many layers of the development process are AI-assisted.
So I think we need to ask:
Who independently verifies the system?
This is not an abstract concern.
Anthropic has said it is using large numbers of AI agents internally and screening their actions, while also discussing external evaluation and transparency around AI-assisted development.
OpenAI has also emphasized monitoring and keeping humans involved as AI research becomes increasingly automated.
The important lesson for me is:
The more AI participates in building AI, the more important independent evaluation becomes.
Not less.
We May Need a New AI Development Architecture
I think the traditional software development lifecycle is going to change.
Today we broadly think in terms of:
Idea → Design → Build → Test → Deploy → Monitor
But an AI-native development process may increasingly look like:
Human Intent → AI Research → AI Design → AI Implementation → AI Experimentation → AI Evaluation → Human Review → Deployment → Monitoring → Feedback → New Research
Notice something important.
The loop is becoming increasingly continuous.
The development process isn’t simply:
Build → Ship.
It becomes:
Hypothesize → Build → Test → Learn → Modify → Test Again.
And AI can participate in multiple stages simultaneously.
That creates enormous potential.
It also creates enormous complexity.
More AI in the Loop Does Not Automatically Mean Better AI
This is where I think we need to be careful.
There is a temptation to look at these developments and conclude:
“If AI can help build AI, then AI development will accelerate indefinitely.”
Maybe.
But that’s not guaranteed.
A development system still has constraints.
You can have unlimited coding capacity and still have limited:
- Compute
- High-quality data
- Evaluation capacity
- Human attention
- Scientific understanding
- Infrastructure
- Energy
- Real-world feedback
And there is another constraint that I consider particularly important:
Verification.
Creating a hypothesis is not the same as proving it.
Generating code is not the same as producing reliable software.
Producing an experiment is not the same as interpreting it correctly.
Generating a more capable model is not the same as demonstrating that the model is safe or reliable.
This is why I think evaluation will become one of the central disciplines of the AI era.
The Human Role May Move From Builder to Governor
I don’t think “AI replaces developers” is the most useful way to frame what is happening.
A better question is:
How does the role of a technical professional change when AI can perform more of the execution?
A developer may spend less time typing code and more time:
- Designing systems
- Defining constraints
- Reviewing architecture
- Evaluating AI output
- Managing agents
- Understanding failure modes
- Making tradeoffs
A researcher may spend less time implementing experiments and more time:
- Selecting research questions
- Designing evaluation methods
- Interpreting results
- Challenging assumptions
- Setting research direction
A manager may spend less time coordinating tasks and more time:
- Setting priorities
- Allocating resources
- Defining boundaries
- Evaluating outcomes
This is a shift from execution toward orchestration and judgment.
And I think that shift will happen across many professions.
This Is Why I Keep Coming Back to Systems
One of the biggest lessons I have taken from working across AI, process improvement and technology is that components rarely tell the whole story.
A model is a component.
An agent is a component.
A prompt is a component.
A tool is a component.
A database is a component.
A human is a component.
But the business outcome comes from the system connecting them.
That is why I recently wrote about moving from building AI agents toward building complete AI systems.
And this new development reinforces that idea.
If AI is increasingly involved in building AI, we shouldn’t only ask:
“How capable is the model?”
We should ask:
“How well-designed is the system in which that model operates?”
That includes:
- Human oversight
- Evaluation
- Tool access
- Data
- Infrastructure
- Governance
- Feedback
- Monitoring
- Decision rights
The model is important.
But the system determines what the model actually does in the world.
The Question I Think We Should Be Asking Now
I don’t think the most useful question is:
“Will AI eventually build itself?”
That question is fascinating, but it can become speculative very quickly.
I think the more useful question is:
“How should humans design the development process when AI is already participating in the development of more capable AI?”
That question is actionable.
It affects companies today.
It affects AI labs.
It affects software engineers.
It affects researchers.
It affects governance.
And eventually, it will affect almost anyone building technology.
My Take
I see this development as neither something to celebrate blindly nor something to fear automatically.
I see it as a systems-design challenge.
If AI makes AI development faster, we need to become better at deciding:
What should be built?
How should it be tested?
Who should verify it?
Where should humans remain in control?
What happens when AI disagrees with its evaluator?
How do we detect failures that our automated systems cannot see?
How do we prevent speed from outrunning our ability to understand what we are building?
These questions matter because the development loop itself is changing.
And whenever the loop changes, the operating model must change with it.
What I Would Do
If I were designing an AI development organization today, I would not simply add more AI agents to every stage.
I would build a system around five principles:
1. Let AI execute aggressively where evaluation is strong.
If a task is measurable and reversible, AI should be able to move quickly.
2. Keep humans heavily involved where judgment matters.
Especially around research direction, high-impact decisions, safety and ambiguous outcomes.
3. Separate generation from evaluation.
The system that creates something should not automatically be the only system deciding that it is good.
4. Make the entire development loop observable.
We should know what AI did, why it did it, what changed, what was tested, and where humans intervened.
5. Treat every acceleration as a potential bottleneck shift.
If AI makes one part of the process ten times faster, immediately ask:
“Where is the bottleneck now?”
That question may become one of the most important questions in AI management.
The Bigger Shift
Perhaps the biggest change is not that AI is becoming capable of building AI.
It is that the boundary between using AI and developing AI is beginning to disappear.
When engineers use AI to write the code that improves AI infrastructure, they are simultaneously:
using AI → developing AI → improving the process used to develop AI.
That creates a feedback loop.
And feedback loops can compound.
We should therefore stop thinking about AI development as a linear pipeline.
It increasingly looks like a system:
Human Intent ↔ AI Execution ↔ Evaluation ↔ Feedback ↔ Improvement
The question is not whether humans disappear from that loop.
The question is:
How intelligently do we design the loop?
That, to me, is the real story.
Final Thought
We spent the first few years of the generative AI era asking:
“What can AI do?”
Then we started asking:
“What can AI automate?”
Now we are beginning to ask:
“What can AI build?”
I think the next question is even more important:
“What happens when AI helps decide, design, build and test the next generation of AI?”
We don’t have a complete answer yet.
And I don’t think we should pretend that we do.
But we should start designing for the possibility.
Because the most important transformation may not be happening inside the AI model.
It may be happening inside the process that creates the AI model.
And if that process is going to become increasingly intelligent, then our responsibility is not simply to make it faster.
Our responsibility is to make the entire system worthy of that speed.
About Author
Jaideep Parashar is the Founder & Director of ReThynk AI Innovation and Research Pvt. Ltd., Six Sigma Black Belt, Lean Expert, AI Strategist, researcher, author, and keynote speaker. His work focuses on combining Artificial Intelligence, Lean Six Sigma, systems thinking, and continuous improvement to help organizations build reliable, scalable, and continuously improving AI-powered operations.
To explore new ideas, read the Newsletter with other 30000 global readers: Newsletter Access!
To explore new technologies and AI advancement: Join Dev.to