We are seeing that the cost of coding and software product development is dropping dramatically. In their early testing with Claude Opus 4.6, Rakuten autonomously closed and assigned issues across a 50-person organization, handling both product and organizational decisions and knowing when to escalate. Anthropic in the meantime built a partially functional C compiler in two weeks, with a team of parallel Claudes.
These various improvements in autonomy, cost efficiency etc. are having important consequences on the way organizations go about developing software products, and on the jobs involved in Software Product Development too.
The Six levels of AI-Assisted Software Product Development
Dan Shapiro, CEO of Glowforge, has categorized AI-assisted software development into six levels. Higher levels are harder to get to, but yield much greater outcomes.
Level 0: Spicy Autocomplete
Start typing. The AI suggests the next few characters or lines to save a few seconds. It’s convenient. But we’re in the same old paradigm. At this level, the human is the “Code Master”. The code unmistakably belongs to the human. This is manual labor.
Level 1: The Coding Intern
The human gives the AI discrete, atomic tasks within a file. They ask it to write a function, add unit tests, clean up code, etc. At this level, the human is a “Task Master”. They write the “important” stuff and use the AI for small, scoped chores.
Level 2: The Junior Developer
The AI has multi-file awareness and tools are “AI-native”. This allows larger tasks to be handed over to the AI. At this level, developers feel a massive flow state and productivity boost.
Level 3: The Developer
The AI acts as the Senior Developer, running multiple tasks and tabs in parallel. At this level, the human almost stops writing code. They review PRs and diffs, spending most of their time ensuring the AI hasn’t veered off course.
Level 4: The Engineering Team
At this level, the human writes a spec, argues with the AI about it, and leaves for 12+ hours while it builds and tests. They effectively become a Product Owner / Manager, focusing on the “what” and “why,” checking whether tests pass and whether the value is there.
Level 5: The Dark Factory
At this level, the AI is a sophisticated system that turns goals and high-level ideas into production-ready software autonomously. Humans are “not welcome” in the implementation loop. Instead, they set the vision and design the feedback loops that allow the system to improve continuously.
Where are we now?
As far as I know, no study has been conducted at this point using these 6 levels to accurately tell where the industry as a whole is. But my sense from talking to people and watching teams both inside and outside Japan, is that most teams are still at level 2 or even below, with a limited but rapidly growing number of teams at level 3. Level 4 is very rare. And level 5 teams are clearly less than 1%.
The next few months will be wild…
The revolution is far from over though. I expect the next few months to be absolutely wild for at least three reasons:
- The core model capabilities are going to get even better. All frontier labs are competing fiercely for attention, user adoption, and confidence from investors. All frontier labs fear being left behind (cf Code Red at OpenAI) and the sense of urgency is at its highest. Also, the AIs are now capable of recursive self-improvement (source: Dario Amodei at the WEF).
- The applications, features and tools that are built on top of those models become more accessible, easier to adopt and use, cheaper, etc. This will make adoption easier and broader.
- All sorts of barriers to adoption are being removed including concerns around privacy and security, slow decision making processes in large organizations, as well as individual skills and mindsets.
I expect intense competition at both the individual and the product/organization level.
Individual level
Spotify’s CEO just claimed that his developers have pretty much stopped writing code. Same thing is being reported at Anthropic. Also, OpenAI has announced that they plan to “dramatically slow down” hiring - precisely because they’re delegating Software Engineering work to AIs. If that’s the future of Software Engineering, then Software Engineers who don’t adapt will become obsolete and their employability will drop significantly.
Product/organization level
The benefits of reaching higher levels in Shapiro’s scale for organizations are huge. We’re not just talking about more features being shipped. We’re also talking about more value being created thanks to shorter, more efficient feedback loops.
Short, efficient feedback loops have always been a massive competitive advantage (Eric Ries, Lean Startup) in product development.
In complex environments, finding Product-Market-Fit requires experimentation and an empirical approach. As you reach levels 4 and 5 of Shapiro’s scale, what you get is better feedback loops. Companies that have that, while their competitors don’t because they’re stuck at level 3, will be much more likely to win.
Tools and Techniques supporting each level
As a team moves up Shapiro’s levels, the tools and the practices change alongside the development team’s focus and mental models. Some practices gain in value. Some practices become obsolete.
Level 0 & 1:
At the Spicy Autocomplete and Coding Intern levels, the focus is on individual efficiency within a manual paradigm. The primary value comes from Standard IDEs and Static Code Analysis. The emphasis here is on Clean Code because the human "Source" still needs to maintain a high degree of readability for manual labor. At level 1, Prompt Engineering and breaking down work into small tasks are the core techniques used to get a few % improvements in productivity from using AI.
Level 2:
As you move to The Junior Developer level, the ROI on Context Management spikes. It is no longer enough for the AI to see a single file; it needs to understand the project’s intent and architecture. Tools like Cursor with codebase indexing become central. The engineer's primary task moves toward Context Engineering. That is, curating the right documentation and file-references so the AI can maintain high-quality outputs without hallucinating.
Level 3:
At the Developer level, the volume of code produced by the AI outpaces traditional human review capacity. Consequently, the emphasis must move toward Specification by Example and a strong Definition of Done (DoD). To prevent the AI from veering off-course, the team must prioritize advanced diffing tools and AI-assisted PR reviews. The value of clear Acceptance Criteria becomes paramount here; if you cannot define the "What" with extreme clarity, the "How" (the code) becomes impossible to manage at scale.
Level 4:
At the Engineering Team level, Agentic Workflows (like Claude Code or Devin) operate independently for hours. This drastically increases the value of automated verification. While traditional TDD (Red-Green-Refactor) is a micro-level design tool for humans, Level 4 requires Test-First Development as a philosophy.
The tests are the only "boss" the agent has while you are away. Therefore, an advanced CI/CD pipeline and a comprehensive test suite act as the primary interface between the human manager and the AI agent. The engineer's job is to write a specification so precise that the test suite can autonomously determine if the agent succeeded or failed.
Level 5:
At the Dark Factory level, the product is not the product. The factory is the product… that builds the product. The implementation loop is entirely autonomous, making comprehensive Telemetry and Monitoring the most critical parts of the stack. Since humans are "out of the loop," the emphasis must be on extreme MTTR (Mean Time to Recovery) and Rollback Procedures. The core discipline here is no longer "development" in the traditional sense, but feedback loop design, using comprehensive analytics to ensure the system’s autonomous evolutions align with the organization’s high-level goals.
Before moving on, take a moment to evaluate your current position and trajectory. Honest inspection is a prerequisite to meaingful adaptation.
- Where are you personally on Shapiro's scale? Where is your team?
- Where are your competitors? Are they still debating "Spicy Autocomplete" while you move toward "Agentic Workflows"? Or is it the other way around?
- Where do you intend to be in 3 months? In 6 months? Given the current rate of model improvement, standing still is equivalent to moving backward.
- What is stopping you from moving up one level today? Do you need better tools (e.g., AI-native IDEs), more adaptable governance (e.g., faster security approvals for agentic tools), or simply dedicated time to learn and rebuild your workflow?
The Bigger Picture: Survival in the "SaaS-pocalypse"
I believe that it is a mistake to assume that every organization will eventually become a "Level 5" Dark Factory. I make a living helping others adapt to change and deal with complexity. So it pains me to say that, but moving up this scale requires more than just investment; it requires a degree of agility that many organizations do not possess and won’t have time to acquire before their market gets disrupted. We are, I believe, more likely to enter a period of severe value concentration.
To illustrate this, look at the evolution of revenue per employee:
- In 1990, IBM generated approximately $69 billion in revenue with a workforce of roughly 370,000 people (~$186,000 per employee) (source)
- In 2025, NVIDIA generated approximately $130 billion in revenue with only 36,000 employees (~$3,611,000 per employee). (source)
Meaning that NVIDIA generates 20x more revenue per employee than an equivalent market leader of the previous generation. I believe that this trend will continue.
The Human Cost of Efficiency
I think we have to be direct about the implications: higher levels of AI autonomy mean we will need fewer people to achieve the same, or greater, results. This doesn't just impact Software Engineers, but also Product Management, Design, QA, etc.
As the "manual labor" of coding, testing, and design is offloaded to agents, what it means to have a "career in tech" is changing. And that leads to a set of difficult but necessary questions:
- How does moving from a "Writer" to a "Reviewer" or "Designer of Feedback Loops" change your day-to-day satisfaction?
- Is a career defined by high-level oversight and agent management still the career you signed up for?
- If the answer is yes, what are you doing today to ensure your skills remain relevant in a world where "writing code" (or “executing tests”, “creating design mockups”, etc.) is no longer the primary value-add? The barrier to entry for writing code and building software is falling to zero, but the bar for building valuable, competitive products is rising. So how about exploring that path?
There is so much more to explore, but things are moving so fast that I'll stop here for today. I am hoping that with this article, I have managed to strike the right balance between sharing arguably bad news and, at the same time, giving you clues about what adapting could look like.
I hope you found this helpful and if you want to continue the conversation, feel free to reach out!